{
 "generated": "2026-11-17",
 "count": 227,
 "labs": [
  {
   "orig_id": "aalto-speech",
   "name": "Aalto Speech Recognition",
   "alias": null,
   "full_name": "Aalto Speech Recognition",
   "type": "academic",
   "founded": null,
   "country": "Finland",
   "city": "Espoo",
   "lat": 60.1867,
   "lon": 24.8275,
   "website": "https://github.com/aalto-speech",
   "huggingface": null,
   "github": "https://github.com/aalto-speech",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Speech recognition group at Aalto University, long-term open-source publisher of the Morfessor morphology tool and AaltoASR toolkit.",
   "tagline": "Aalto University's speech recognition group, whose open Morfessor and AaltoASR tools have long supported speech and language technology for morphologically rich languages.",
   "about": [
    "The speech recognition group sits in the Department of Information and Communications Engineering at Aalto University in Espoo and is part of the Finnish Center for Artificial Intelligence. It works on acoustic modeling of phonemes, statistical language models for large-vocabulary continuous speech recognition, and speech technology for under-resourced languages.",
    "The group publishes its research software through the aalto-speech GitHub organization, created in 2013. The best known repositories are Morfessor, a tool for unsupervised and semi-supervised morphological segmentation, AaltoASR, a set of recognition tools, and speaker diarization scripts built on that toolkit.",
    "Professor Mikko Kurimo heads the group and is known internationally for unsupervised subword modeling for languages with rich word inflection, such as Finnish, Estonian, Turkish and Arabic. The software mainly serves other speech and language researchers."
   ],
   "mission": "The group aims to improve speech recognition with new algorithms and to make recognition practical for languages with limited data and heavy inflection. It shares its tools openly so that other researchers can reuse and extend them.",
   "notice": null,
   "known_for": [
    "Morfessor morphological segmentation",
    "AaltoASR recognition toolkit",
    "Subword modeling for inflected languages",
    "Open speaker diarization scripts"
   ],
   "focus": [
    "speech",
    "morphology"
   ],
   "contributions": [
    {
     "title": "Morfessor repository",
     "note": "The Morfessor tool for unsupervised and semi-supervised morphological segmentation appeared on the group's GitHub organization under a BSD licence.",
     "year": "2013"
    },
    {
     "title": "AaltoASR toolkit",
     "note": "The group published the Aalto Automatic Speech Recognition tools as an open repository.",
     "year": "2013"
    },
    {
     "title": "Speaker diarization scripts",
     "note": "Diarization scripts based on AaltoASR were released and have drawn close to 200 GitHub stars.",
     "year": "2016"
    },
    {
     "title": "Subword Kaldi recipes",
     "note": "A repository showing how to handle position-dependent phones in a subword lexicon FST for Kaldi was published.",
     "year": "2018"
    }
   ],
   "github_highlights": [
    {
     "name": "morfessor",
     "url": "https://github.com/aalto-speech/morfessor",
     "note": "Morfessor is a tool for unsupervised and semi-supervised morphological segmentation (207 stars)"
    },
    {
     "name": "speaker-diarization",
     "url": "https://github.com/aalto-speech/speaker-diarization",
     "note": "Speaker diarization scripts, based on AaltoASR (192 stars)"
    },
    {
     "name": "AaltoASR",
     "url": "https://github.com/aalto-speech/AaltoASR",
     "note": "Aalto Automatic Speech Recognition tools (90 stars)"
    },
    {
     "name": "subword-kaldi",
     "url": "https://github.com/aalto-speech/subword-kaldi",
     "note": "Properly handle position-dependent phones in a subword lexicon FST (31 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Morfessor",
     "url": "https://github.com/aalto-speech/morfessor",
     "note": "Unsupervised morphological segmentation tool"
    },
    {
     "name": "speaker-diarization",
     "url": "https://github.com/aalto-speech/speaker-diarization",
     "note": "Diarization scripts based on AaltoASR"
    },
    {
     "name": "AaltoASR",
     "url": "https://github.com/aalto-speech/AaltoASR",
     "note": "Aalto automatic speech recognition tools"
    },
    {
     "name": "subword-kaldi",
     "url": "https://github.com/aalto-speech/subword-kaldi",
     "note": "Subword lexicon FST handling for Kaldi"
    }
   ],
   "openness": "Tools are open source on GitHub.",
   "key_people": [
    {
     "name": "Mikko Kurimo",
     "role": "Professor, head of the speech recognition group",
     "url": "https://www.aalto.fi/en/people/mikko-kurimo",
     "note": null
    },
    {
     "name": "Peter Smit",
     "role": "Public member of the aalto-speech GitHub organization",
     "url": "http://petersmit.eu",
     "note": null
    },
    {
     "name": "Mittul Singh",
     "role": "Research Fellow",
     "url": "https://www.aalto.fi/en/department-of-signal-processing-and-acoustics/speech-recognition",
     "note": null
    },
    {
     "name": "Anssi Moisio",
     "role": "Doctoral researcher",
     "url": "https://www.aalto.fi/en/department-of-signal-processing-and-acoustics/speech-recognition",
     "note": null
    },
    {
     "name": "Yaroslav Getman",
     "role": "Doctoral researcher",
     "url": "https://www.aalto.fi/en/department-of-signal-processing-and-acoustics/speech-recognition",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://github.com/aalto-speech",
    "https://github.com/aalto-speech"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "aalto-speech-recognition",
   "logo": true
  },
  {
   "orig_id": "adapterhub",
   "name": "AdapterHub",
   "alias": null,
   "full_name": "AdapterHub",
   "type": "academic",
   "founded": 2020,
   "country": "Germany",
   "city": "Darmstadt",
   "lat": 49.8728,
   "lon": 8.6512,
   "website": "https://adapterhub.ml",
   "huggingface": "https://huggingface.co/AdapterHub",
   "github": "https://github.com/adapter-hub",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source community and library for parameter-efficient and modular transfer learning, started by researchers around UKP Lab in Darmstadt.",
   "tagline": "AdapterHub maintains the open Adapters library, a modular, parameter-efficient fine-tuning toolkit built on Hugging Face Transformers.",
   "about": [
    "AdapterHub began in 2020 as a framework for sharing and composing adapters, small sets of added weights that stand in for full fine-tuning of a pretrained model. The original paper by Pfeiffer and colleagues was presented at EMNLP 2020, and the GitHub organization was created in April 2020.",
    "In November 2023 the project released the unified Adapters package, replacing the earlier adapter-transformers fork, with a library paper at EMNLP 2023. The library is released under the Apache 2.0 licence and had about 2,800 GitHub stars in 2026.",
    "Version 1.2.0, announced in May 2025, added a plugin interface for attaching adapter functionality to nearly any Transformer model. The project is run by researchers associated with TU Darmstadt and serves practitioners who need to adapt large models cheaply."
   ],
   "mission": "AdapterHub aims to make parameter-efficient and modular transfer learning easy to use, share and combine. It treats adapters as reusable components that the community can publish and compose.",
   "notice": null,
   "known_for": [
    "Adapters library for Transformers",
    "Parameter-efficient fine-tuning",
    "Sharing and composing adapter modules",
    "Origin in adapter-transformers"
   ],
   "focus": [
    "parameter-efficient fine-tuning",
    "modular learning",
    "NLP"
   ],
   "contributions": [
    {
     "title": "AdapterHub framework",
     "note": "Pfeiffer et al. introduced a framework for sharing and composing adapters, presented at EMNLP 2020.",
     "year": "2020"
    },
    {
     "title": "adapter-transformers",
     "note": "The repository for the original adapter extension of Hugging Face Transformers was created in April 2020.",
     "year": "2020"
    },
    {
     "title": "Adapters library paper",
     "note": "The unified Adapters package and its library paper were released and presented at EMNLP 2023.",
     "year": "2023"
    },
    {
     "title": "Adapters v1.2.0",
     "note": "A plugin interface was added so adapter functionality can be attached to nearly any Transformer model.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "adapters",
     "url": "https://github.com/adapter-hub/adapters",
     "note": "A Unified Library for Parameter-Efficient and Modular Transfer Learning  (2,833 stars)"
    },
    {
     "name": "Hub",
     "url": "https://github.com/adapter-hub/Hub",
     "note": "ARCHIVED. Please use https://docs.adapterhub.ml/huggingface_hub.html || 🔌 A central repository collecting pre- (69 stars)"
    },
    {
     "name": "efficient-task-transfer",
     "url": "https://github.com/adapter-hub/efficient-task-transfer",
     "note": "Research code for \"What to Pre-Train on? Efficient Intermediate Task Selection\", EMNLP 2021 (37 stars)"
    },
    {
     "name": "hgiyt",
     "url": "https://github.com/adapter-hub/hgiyt",
     "note": "Research code for the paper \"How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Langua (27 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "AdapterHub/roberta-base-pf-scitail",
     "url": "https://huggingface.co/AdapterHub/roberta-base-pf-scitail",
     "kind": "model",
     "note": "Model with 9 likes on the Hub."
    },
    {
     "name": "AdapterHub/bert-base-uncased-pf-conll2000",
     "url": "https://huggingface.co/AdapterHub/bert-base-uncased-pf-conll2000",
     "kind": "model",
     "note": "Model with 8 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "AdapterHub: A Framework for Adapting Transformers",
     "url": "https://arxiv.org/abs/2007.07779",
     "year": "2020"
    },
    {
     "title": "Adapters: A Unified Library for Parameter-Efficient and Modular Transfer Learning",
     "url": "https://arxiv.org/abs/2311.11077",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "Adapters library",
     "url": "https://github.com/adapter-hub/adapters",
     "note": "Unified library for parameter-efficient and modular transfer learning"
    },
    {
     "name": "AdapterHub Hub",
     "url": "https://github.com/adapter-hub/Hub",
     "note": "Archived central repository of pre-trained adapters"
    },
    {
     "name": "Adapter checkpoints",
     "url": "https://huggingface.co/AdapterHub/roberta-base-pf-scitail",
     "note": "Pre-trained adapter on the Hub"
    }
   ],
   "openness": "Fully open-source library and adapter checkpoints on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Jonas Pfeiffer",
     "role": "Lead author of the original AdapterHub paper and library paper",
     "url": "https://github.com/JoPfeiffer",
     "note": null
    },
    {
     "name": "Clifton Poth",
     "role": "First author of the Adapters library paper and maintainer",
     "url": "https://github.com/calpt",
     "note": null
    },
    {
     "name": "Hannah Sterz",
     "role": "Author of the Adapters library paper",
     "url": "https://github.com/hSterz",
     "note": null
    },
    {
     "name": "Leon Engländer",
     "role": "Author of the Adapters library paper",
     "url": "https://github.com/lenglaender",
     "note": null
    },
    {
     "name": "Andreas Rücklé",
     "role": "Member of the adapter-hub GitHub organization",
     "url": "http://rueckle.net",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the project website.",
   "sources": [
    "https://adapterhub.ml",
    "https://huggingface.co/AdapterHub",
    "https://github.com/adapter-hub",
    "https://arxiv.org/abs/2007.07779"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "adapterhub",
   "mlat": 49.8295,
   "mlon": 8.6512,
   "logo": true
  },
  {
   "orig_id": "agentica",
   "name": "Agentica",
   "alias": "Berkeley Sky Lab",
   "full_name": "Agentica (Berkeley Sky Lab)",
   "type": "academic",
   "founded": null,
   "country": "United States",
   "city": "Berkeley",
   "lat": 37.8715,
   "lon": -122.273,
   "website": "https://agentica-project.com",
   "huggingface": "https://huggingface.co/agentica-org",
   "github": "https://github.com/agentica-project",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source initiative from Berkeley's Sky Computing Lab releasing RL-trained models such as DeepScaleR and DeepCoder and the rLLM framework.",
   "tagline": "A Berkeley Sky Computing Lab initiative that publishes open reinforcement learning recipes, models and the rLLM framework for training language agents.",
   "about": [
    "Agentica is an open-source effort from the Sky Computing Lab at UC Berkeley. It aims to make reinforcement learning techniques for language models and agents widely available, with core contributors Michael Luo and Sijun Tan and a group of student contributors.",
    "The project released DeepScaleR in February 2025, a 1.5B model trained with distributed reinforcement learning for math reasoning. In July 2025 it published DeepCoder, a 14B code reasoning model, the rLLM framework for training agents with reinforcement learning, and DeepSWE, a 32B software engineering agent.",
    "Its advisors are Li Erran Li, Raluca Ada Popa and Ion Stoica. The project site notes that Agentica is becoming a startup named Peano AI."
   ],
   "mission": "Agentica states that it wants to democratize reinforcement learning for language models and to build scalable systems for LLMs and agents.",
   "notice": null,
   "known_for": [
    "DeepScaleR reinforcement-learning math model",
    "DeepCoder code reasoning model",
    "The rLLM agent training framework",
    "DeepSWE software engineering agent"
   ],
   "focus": [
    "reinforcement learning",
    "reasoning",
    "coding"
   ],
   "contributions": [
    {
     "title": "DeepScaleR released",
     "note": "A 1.5B model trained with distributed reinforcement learning.",
     "year": "2025"
    },
    {
     "title": "DeepCoder released",
     "note": "A 14B open code reasoning model.",
     "year": "2025"
    },
    {
     "title": "rLLM released",
     "note": "An open framework for training language agents with reinforcement learning.",
     "year": "2025"
    },
    {
     "title": "DeepSWE released",
     "note": "A 32B software engineering agent trained with reinforcement learning.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "rllm",
     "url": "https://github.com/agentica-project/rllm",
     "note": "RL for language agents."
    },
    {
     "name": "verl-pipeline",
     "url": "https://github.com/agentica-project/verl-pipeline",
     "note": "Async pipelined verl."
    },
    {
     "name": "R2E-Gym",
     "url": "https://github.com/agentica-project/R2E-Gym",
     "note": "R2E-Gym."
    }
   ],
   "hf_highlights": [
    {
     "name": "DeepScaleR-1.5B-Preview",
     "url": "https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview",
     "kind": "model",
     "note": "DeepScaleR."
    },
    {
     "name": "DeepCoder-14B-Preview",
     "url": "https://huggingface.co/agentica-org/DeepCoder-14B-Preview",
     "kind": "model",
     "note": "DeepCoder."
    },
    {
     "name": "DeepScaleR-Preview-Dataset",
     "url": "https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset",
     "kind": "dataset",
     "note": "Training data."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "rllm",
     "url": "https://github.com/agentica-project/rllm",
     "note": "RL framework"
    },
    {
     "name": "DeepCoder-14B-Preview",
     "url": "https://huggingface.co/agentica-org/DeepCoder-14B-Preview",
     "note": "Model"
    }
   ],
   "openness": "Open weights, data and code.",
   "key_people": [
    {
     "name": "Michael Luo",
     "role": "Core contributor",
     "url": "https://agentica-project.com",
     "note": null
    },
    {
     "name": "Sijun Tan",
     "role": "Core contributor",
     "url": "https://agentica-project.com",
     "note": null
    },
    {
     "name": "Ion Stoica",
     "role": "Advisor",
     "url": "https://agentica-project.com",
     "note": "Listed as advisor on the project site."
    },
    {
     "name": "Raluca Ada Popa",
     "role": "Advisor",
     "url": "https://agentica-project.com",
     "note": null
    },
    {
     "name": "Li Erran Li",
     "role": "Advisor",
     "url": "https://agentica-project.com",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://agentica-project.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://agentica-project.com",
    "https://huggingface.co/agentica-org"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "agentica",
   "mlat": 37.8156,
   "mlon": -122.273,
   "logo": true
  },
  {
   "orig_id": "ai-singapore",
   "name": "AI Singapore",
   "alias": "SEA-LION",
   "full_name": "AI Singapore (SEA-LION)",
   "type": "nonprofit",
   "founded": 2017,
   "country": "Singapore",
   "city": "Singapore",
   "lat": 1.35,
   "lon": 103.82,
   "website": "https://www.aisingapore.org",
   "huggingface": "https://huggingface.co/aisingapore",
   "github": "https://github.com/aisingapore",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "National AI programme (government-funded) behind the SEA-LION family of Southeast Asian language models.",
   "tagline": "Singapore's national AI programme, which builds applied AI capability and the open SEA-LION language models for Southeast Asia.",
   "about": [
    "AI Singapore was launched in May 2017 as a national initiative that brings together Singapore's research institutions, start-ups and companies. Its work covers applied research, AI governance, innovation support for companies, talent development and open-source AI products.",
    "Among its products is SEA-LION, a family of open language models for Southeast Asian languages and cultures. Reporting describes a first release in 2023 trained on 11 languages, a version 3 built on Llama and Gemma 2, and a version 4 built on Gemma 3 and on Qwen3 with technical support from Alibaba.",
    "The programme also runs talent schemes such as the AI Apprenticeship Programme and offers courses through LearnAI."
   ],
   "mission": "AI Singapore states that it aims to strengthen national AI capabilities, build tools and train talent, and to raise Singapore's profile in AI.",
   "notice": null,
   "known_for": [
    "SEA-LION Southeast Asian language models",
    "National AI apprenticeship and talent programmes",
    "Government-backed applied AI research"
   ],
   "focus": [
    "language models",
    "Southeast Asian languages"
   ],
   "contributions": [
    {
     "title": "Programme launched",
     "note": "AI Singapore was established as a national AI initiative.",
     "year": "2017"
    },
    {
     "title": "SEA-LION first release",
     "note": "The first open model in the family, trained on 11 Southeast Asian languages.",
     "year": "2023"
    },
    {
     "title": "SEA-LION technical report",
     "note": "The team published an arXiv report describing the model family.",
     "year": "2025"
    },
    {
     "title": "SEA-LION v4 released",
     "note": "Gemma-based v4 models were published, with a Qwen-based line following.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "TagUI",
     "url": "https://github.com/aisingapore/TagUI",
     "note": "Free RPA tool, the org's most-starred repo."
    },
    {
     "name": "sealion",
     "url": "https://github.com/aisingapore/sealion",
     "note": "Entry repo for SEA-LION models."
    },
    {
     "name": "PeekingDuck",
     "url": "https://github.com/aisingapore/PeekingDuck",
     "note": "Modular computer vision inference framework."
    },
    {
     "name": "TagUI-Bricks",
     "url": "https://github.com/aisingapore/TagUI-Bricks",
     "note": "Collection of TagUI workflows."
    }
   ],
   "hf_highlights": [
    {
     "name": "aisingapore/Apertus-SEA-LION-v4-8B-IT-GGUF",
     "url": "https://huggingface.co/aisingapore/Apertus-SEA-LION-v4-8B-IT-GGUF",
     "kind": "model",
     "note": "SEA-LION v4 8B instruct, GGUF."
    },
    {
     "name": "aisingapore/Gemma-SEA-LION-v3-9B-IT",
     "url": "https://huggingface.co/aisingapore/Gemma-SEA-LION-v3-9B-IT",
     "kind": "model",
     "note": "Gemma-based SEA-LION v3."
    },
    {
     "name": "aisingapore/NLG-Machine-Translation",
     "url": "https://huggingface.co/datasets/aisingapore/NLG-Machine-Translation",
     "kind": "dataset",
     "note": "Regional translation evaluation data."
    },
    {
     "name": "aisingapore/NLU-Sentiment-Analysis",
     "url": "https://huggingface.co/datasets/aisingapore/NLU-Sentiment-Analysis",
     "kind": "dataset",
     "note": "Sentiment evaluation data."
    }
   ],
   "papers": [
    {
     "title": "SEA-LION: Southeast Asian Languages in One Network",
     "url": "https://arxiv.org/abs/2504.05747",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Gemma-SEA-LION-v4-27B-IT",
     "url": "https://huggingface.co/aisingapore/Gemma-SEA-LION-v4-27B-IT",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Gemma-SEA-LION-v4-27B-VL",
     "url": "https://huggingface.co/aisingapore/Gemma-SEA-LION-v4-27B-VL",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face and code on GitHub.",
   "key_people": [
    {
     "name": "Leslie Teo",
     "role": "Senior Director, AI Products; SEA-LION lead",
     "url": "https://arxiv.org/abs/2504.05747",
     "note": "Co-author of the SEA-LION technical report."
    },
    {
     "name": "Mohan Kankanhalli",
     "role": "Deputy Executive Chairman; Director, NUS AI Institute",
     "url": "https://www.comp.nus.edu.sg/~mohan/",
     "note": null
    },
    {
     "name": "Simon Chesterman",
     "role": "Senior Director, AI Governance",
     "url": "https://en.wikipedia.org/wiki/Simon_Chesterman",
     "note": "Also a professor at the National University of Singapore."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact and programme details are on aisingapore.org; SEA-LION materials are on GitHub and Hugging Face.",
   "sources": [
    "https://huggingface.co/aisingapore",
    "https://www.aisingapore.org"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "ai-singapore",
   "mlat": 1.31464,
   "mlon": 103.82,
   "logo": true
  },
  {
   "orig_id": "ai-sweden",
   "name": "AI Sweden",
   "alias": null,
   "full_name": "AI Sweden",
   "type": "institute",
   "founded": 2019,
   "country": "Sweden",
   "city": "Gothenburg",
   "lat": 57.7089,
   "lon": 11.9746,
   "website": "https://www.ai.se",
   "huggingface": "https://huggingface.co/AI-Sweden-Models",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/aisweden",
   "contact_email": null,
   "summary": "Sweden's national centre for applied AI, a partnership between public and private organisations. Its model hub hosts the GPT-SW3 Nordic language models and related resources.",
   "tagline": "Sweden's national centre for applied AI, a public and private partnership that also published the GPT-SW3 Nordic language models.",
   "about": [
    "AI Sweden describes itself as a national centre for applied AI. It is not-for-profit and works with more than 180 partners from companies, public bodies and universities.",
    "Its programmes cover AI adoption and training, sector initiatives such as healthcare and automotive, a startup programme and research on infrastructure, secure AI and language technology.",
    "In language technology it developed GPT-SW3 with RISE and WASP, a family of autoregressive models for Swedish and the other Nordic languages, with model cards dated December 2022 and an open release on Hugging Face reported in late 2023."
   ],
   "mission": "AI Sweden states that its purpose is to accelerate the use of AI for the benefit of society.",
   "notice": null,
   "known_for": [
    "GPT-SW3 Nordic language models",
    "National applied AI partnership",
    "Shared datasets and testbeds for Swedish AI"
   ],
   "focus": [
    "language models",
    "Swedish",
    "Nordic languages",
    "applied AI"
   ],
   "contributions": [
    {
     "title": "GPT-SW3 models documented",
     "note": "Hugging Face model cards for the first GPT-SW3 sizes are dated December 2022.",
     "year": "2022"
    },
    {
     "title": "GPT-SW3 paper",
     "note": "Technical description of the Nordic autoregressive models, posted to arXiv in May 2023.",
     "year": "2023"
    },
    {
     "title": "GPT-SW3 open release",
     "note": "Models from 126M to 40B parameters were made available on Hugging Face under a permissive licence.",
     "year": "2023"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "AI-Sweden-Models/Llama-3-8B-instruct",
     "url": "https://huggingface.co/AI-Sweden-Models/Llama-3-8B-instruct",
     "kind": "model",
     "note": "Model: 1,470 downloads, 13 likes."
    },
    {
     "name": "AI-Sweden-Models/roberta-large-1160k",
     "url": "https://huggingface.co/AI-Sweden-Models/roberta-large-1160k",
     "kind": "model",
     "note": "Model: 694 downloads, 12 likes."
    },
    {
     "name": "AI-Sweden-Models/BiaSWE",
     "url": "https://huggingface.co/datasets/AI-Sweden-Models/BiaSWE",
     "kind": "dataset",
     "note": "Dataset: 134 downloads, 4 likes."
    },
    {
     "name": "AI-Sweden-Models/ModernBERT-large",
     "url": "https://huggingface.co/AI-Sweden-Models/ModernBERT-large",
     "kind": "model",
     "note": "Model: 499 downloads, 7 likes."
    }
   ],
   "papers": [
    {
     "title": "GPT-SW3: An Autoregressive Language Model for the Nordic Languages",
     "url": "https://arxiv.org/abs/2305.12987",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "GPT-SW3 40B",
     "url": "https://huggingface.co/AI-Sweden-Models/gpt-sw3-40b",
     "note": "Nordic-focused GPT model"
    },
    {
     "name": "GPT-SW3 6.7B v2 Instruct",
     "url": "https://huggingface.co/AI-Sweden-Models/gpt-sw3-6.7b-v2-instruct",
     "note": "Instruction-tuned variant"
    },
    {
     "name": "BiaSWE",
     "url": "https://huggingface.co/datasets/AI-Sweden-Models/BiaSWE",
     "note": "Swedish bias evaluation dataset"
    }
   ],
   "openness": "Publishes models and datasets on Hugging Face; individual licences vary, so check each model card.",
   "key_people": [
    {
     "name": "Magnus Sahlgren",
     "role": "Researcher, GPT-SW3 co-author",
     "url": "https://arxiv.org/abs/2305.12987",
     "note": null
    },
    {
     "name": "Ariel Ekgren",
     "role": "Lead author, GPT-SW3",
     "url": "https://arxiv.org/abs/2305.12987",
     "note": null
    },
    {
     "name": "Joey Öhman",
     "role": "Researcher, GPT-SW3 co-author",
     "url": "https://arxiv.org/abs/2305.12987",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website; staff addresses are listed individually so none is given here.",
   "sources": [
    "https://www.ai.se",
    "https://huggingface.co/AI-Sweden-Models"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "ai-sweden",
   "mlat": 57.67354,
   "mlon": 11.9746,
   "logo": true
  },
  {
   "orig_id": "ai21-labs",
   "name": "AI21 Labs",
   "alias": null,
   "full_name": "AI21 Labs",
   "type": "startup",
   "founded": 2017,
   "country": "Israel",
   "city": "Tel Aviv",
   "lat": 32.09,
   "lon": 34.78,
   "website": "https://ai21.com",
   "huggingface": "https://huggingface.co/ai21labs",
   "github": "https://github.com/AI21Labs",
   "x": "https://twitter.com/AI21Labs",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Israeli enterprise AI company. Partly open: Jamba models are published on Hugging Face while products are commercial.",
   "tagline": "An Israeli enterprise AI company that publishes the hybrid SSM-Transformer Jamba models as open weights alongside commercial products.",
   "about": [
    "AI21 Labs was founded in Tel Aviv in November 2017 by Yoav Shoham, Ori Goshen and Amnon Shashua. Shoham and Goshen serve as co-CEOs and Shashua is chairman.",
    "The company launched the Jurassic-1 language models in August 2021 and released the first Jamba model in March 2024. Jamba 1.5 followed in 2024 and Jamba 1.6 in March 2025, and its Hugging Face organisation now lists Jamba2 and Jamba Reasoning models.",
    "Its main business is commercial AI for enterprises, with the open Jamba weights published alongside paid products."
   ],
   "mission": "AI21 describes its aim as building reliable, practical and scalable AI for the enterprise.",
   "notice": "AI21 is a commercial company. Only some of its Jamba models are released as open weights.",
   "known_for": [
    "Jamba hybrid Mamba-Transformer models",
    "Jurassic-1 language models",
    "Enterprise language AI"
   ],
   "focus": [
    "language models",
    "enterprise"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "AI21 Labs started in Tel Aviv.",
     "year": "2017"
    },
    {
     "title": "Jurassic-1 launched",
     "note": "A language model with a vocabulary of more than 250,000 tokens.",
     "year": "2021"
    },
    {
     "title": "Jamba released",
     "note": "The first Jamba model, a hybrid state-space and Transformer model published as open weights.",
     "year": "2024"
    },
    {
     "title": "Jamba 1.6 released",
     "note": "An updated model family aimed at private enterprise deployment.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "in-context-ralm",
     "url": "https://github.com/AI21Labs/in-context-ralm",
     "note": "In-context retrieval-augmented LMs."
    },
    {
     "name": "lm-evaluation",
     "url": "https://github.com/AI21Labs/lm-evaluation",
     "note": "LLM evaluation suite."
    },
    {
     "name": "Parallel-Context-Windows",
     "url": "https://github.com/AI21Labs/Parallel-Context-Windows",
     "note": "Context window extension."
    },
    {
     "name": "ai21-python",
     "url": "https://github.com/AI21Labs/ai21-python",
     "note": "Python SDK."
    }
   ],
   "hf_highlights": [
    {
     "name": "ai21labs/Jamba-tiny-dev",
     "url": "https://huggingface.co/ai21labs/Jamba-tiny-dev",
     "kind": "model",
     "note": "Small dev Jamba."
    },
    {
     "name": "ai21labs/AI21-Jamba-Reasoning-3B",
     "url": "https://huggingface.co/ai21labs/AI21-Jamba-Reasoning-3B",
     "kind": "model",
     "note": "3B reasoning model."
    },
    {
     "name": "ai21labs/AI21-Jamba2-3B",
     "url": "https://huggingface.co/ai21labs/AI21-Jamba2-3B",
     "kind": "model",
     "note": "Jamba2 3B."
    },
    {
     "name": "ai21labs/LIFTBench",
     "url": "https://huggingface.co/datasets/ai21labs/LIFTBench",
     "kind": "dataset",
     "note": "Benchmark."
    }
   ],
   "papers": [
    {
     "title": "Jamba: A Hybrid Transformer-Mamba Language Model",
     "url": "https://arxiv.org/abs/2403.19887",
     "year": "2024"
    },
    {
     "title": "Jamba-1.5: Hybrid Transformer-Mamba Models at Scale",
     "url": "https://arxiv.org/abs/2408.12570",
     "year": "2024"
    },
    {
     "title": "In-Context Retrieval-Augmented Language Models",
     "url": "https://arxiv.org/abs/2302.00083",
     "year": "2023"
    },
    {
     "title": "Parallel Context Windows for Large Language Models",
     "url": "https://arxiv.org/abs/2212.10947",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "AI21-Jamba2-Mini",
     "url": "https://huggingface.co/ai21labs/AI21-Jamba2-Mini",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "AI21-Jamba2-3B",
     "url": "https://huggingface.co/ai21labs/AI21-Jamba2-3B",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Selected open weights; licence per model card.",
   "key_people": [
    {
     "name": "Yoav Shoham",
     "role": "Co-founder and co-CEO",
     "url": "https://en.wikipedia.org/wiki/Yoav_Shoham",
     "note": null
    },
    {
     "name": "Ori Goshen",
     "role": "Co-founder and co-CEO",
     "url": "https://www.ai21.com/",
     "note": null
    },
    {
     "name": "Amnon Shashua",
     "role": "Co-founder and chairman",
     "url": "https://en.wikipedia.org/wiki/Amnon_Shashua",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "ai21.com has contact and sales forms; open work is on GitHub and Hugging Face.",
   "sources": [
    "https://huggingface.co/ai21labs",
    "https://ai21.com"
   ],
   "confidence": "high",
   "region": "Middle East",
   "id": "ai21-labs",
   "logo": true
  },
  {
   "orig_id": "ai4bharat",
   "name": "AI4Bharat",
   "alias": null,
   "full_name": "AI4Bharat",
   "type": "nonprofit",
   "founded": null,
   "country": "India",
   "city": "Chennai",
   "lat": 13.08,
   "lon": 80.27,
   "website": "https://ai4bharat.iitm.ac.in/",
   "huggingface": "https://huggingface.co/ai4bharat",
   "github": "https://github.com/AI4Bharat",
   "x": "https://twitter.com/ai4bharat",
   "discord": "https://discord.gg/HqJXnJNDaq",
   "linkedin": null,
   "contact_email": null,
   "summary": "Non-profit initiative at IIT Madras building open datasets, models and applications for Indian languages.",
   "tagline": "An IIT Madras research lab building open datasets, models and tools for Indian languages in translation, speech and language modelling.",
   "about": [
    "AI4Bharat is a research lab at IIT Madras that releases most of its work as open source. Its focus areas are transliteration, language understanding and generation, machine translation, speech recognition and speech synthesis.",
    "Its releases include IndicBERT and IndicBART, the IndicTrans2 translation models covering all 22 scheduled Indian languages and English, speech models such as IndicWav2Vec and IndicWhisper, and datasets including Sangraha and IndicVoices.",
    "IndicVoices, funded by the Bhashini initiative of India's IT ministry, provides about 7,300 hours of speech across 22 languages."
   ],
   "mission": "AI4Bharat works to give speakers of Indian languages open language technology comparable to what exists for English.",
   "notice": null,
   "known_for": [
    "IndicTrans2 translation for 22 Indian languages",
    "IndicVoices speech dataset",
    "Open datasets such as Sangraha"
   ],
   "focus": [
    "datasets",
    "speech",
    "translation",
    "Indic languages"
   ],
   "contributions": [
    {
     "title": "Nilekani Centre launched",
     "note": "IIT Madras launched the Nilekani Centre at AI4Bharat in July 2022, funded by a Rs 36 crore grant from Nilekani Philanthropies.",
     "year": "2022"
    },
    {
     "title": "IndicTrans2 paper published",
     "note": "Open translation models for the 22 scheduled Indian languages, with permissive licences.",
     "year": "2023"
    },
    {
     "title": "IndicVoices released",
     "note": "About 7,300 hours of speech from over 16,000 speakers, used to build IndicASR.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "indicnlp_catalog",
     "url": "https://github.com/AI4Bharat/indicnlp_catalog",
     "note": "Catalog of NLP resources for Indic languages."
    },
    {
     "name": "IndicTrans2",
     "url": "https://github.com/AI4Bharat/IndicTrans2",
     "note": "Translation models for 22 languages."
    },
    {
     "name": "Indic-TTS",
     "url": "https://github.com/AI4Bharat/Indic-TTS",
     "note": "Text-to-speech for Indian languages."
    },
    {
     "name": "IndicLLMSuite",
     "url": "https://github.com/AI4Bharat/IndicLLMSuite",
     "note": "Blueprint for Indic pretraining datasets."
    }
   ],
   "hf_highlights": [
    {
     "name": "ai4bharat/indictrans2-en-indic-dist-200M",
     "url": "https://huggingface.co/ai4bharat/indictrans2-en-indic-dist-200M",
     "kind": "model",
     "note": "Distilled translation model."
    },
    {
     "name": "ai4bharat/IndicBART",
     "url": "https://huggingface.co/ai4bharat/IndicBART",
     "kind": "model",
     "note": "Seq2seq model for Indic languages."
    },
    {
     "name": "ai4bharat/indic-parler-tts",
     "url": "https://huggingface.co/ai4bharat/indic-parler-tts",
     "kind": "model",
     "note": "Indic TTS."
    },
    {
     "name": "ai4bharat/sangraha",
     "url": "https://huggingface.co/datasets/ai4bharat/sangraha",
     "kind": "dataset",
     "note": "Indic pretraining corpus."
    }
   ],
   "papers": [
    {
     "title": "IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages",
     "url": "https://arxiv.org/abs/2305.16307",
     "year": "2023"
    },
    {
     "title": "Towards Leaving No Indic Language Behind: Building Monolingual Corpora, Benchmark and Models for Indic Languages",
     "url": "https://arxiv.org/abs/2212.05409",
     "year": "2022"
    },
    {
     "title": "IndicLLMSuite: A Blueprint for Creating Pre-training and Fine-Tuning Datasets for Indian Languages",
     "url": "https://arxiv.org/abs/2403.06350",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "indic-conformer-600m-multilingual",
     "url": "https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "indic-seamless",
     "url": "https://huggingface.co/ai4bharat/indic-seamless",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "indictrans2-indic-en-dist-200M",
     "url": "https://huggingface.co/ai4bharat/indictrans2-indic-en-dist-200M",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open datasets, models and code.",
   "key_people": [
    {
     "name": "Mitesh M. Khapra",
     "role": "Professor, IIT Madras; head of AI4Bharat",
     "url": "https://www.cse.iitm.ac.in/~miteshk/",
     "note": "Co-founder of the lab."
    },
    {
     "name": "Pratyush Kumar",
     "role": "Co-founder; adjunct faculty, IIT Madras",
     "url": "https://arxiv.org/abs/2305.16307",
     "note": "IndicTrans2 co-author."
    },
    {
     "name": "Anoop Kunchukuttan",
     "role": "Founding member and co-lead; Microsoft researcher",
     "url": "https://anoopkunchukuttan.gitlab.io/",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the AI4Bharat site.",
   "contact_note": "AI4Bharat's website links a community Discord and its GitHub issue trackers.",
   "sources": [
    "https://huggingface.co/ai4bharat",
    "https://ai4bharat.iitm.ac.in/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "ai4bharat",
   "logo": true
  },
  {
   "orig_id": "lumii-ailab",
   "name": "AiLab",
   "alias": "University of Latvia IMCS",
   "full_name": "AiLab (University of Latvia IMCS)",
   "type": "academic",
   "founded": 1992,
   "country": "Latvia",
   "city": "Riga",
   "lat": 56.9496,
   "lon": 24.1052,
   "website": "https://ailab.lv",
   "huggingface": null,
   "github": "https://github.com/LUMII-AILab",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@ailab.lv",
   "summary": "Artificial Intelligence Laboratory at the Institute of Mathematics and Computer Science, University of Latvia, active since 1992 and building open Latvian language resources such as Tezaurs and FullStack-LV.",
   "tagline": "The AI Laboratory at the University of Latvia builds open Latvian language resources, from the Tezaurs lexicon to a multilayer corpus and Latvian BERT models.",
   "about": [
    "The Artificial Intelligence Laboratory belongs to the Institute of Mathematics and Computer Science at the University of Latvia in Riga and has conducted research since 1992. Its work covers natural language processing and machine learning, including Latvian speech recognition and synthesis, syntactic and semantic analysis and language models.",
    "The lab maintains a set of public Latvian resources, among them Tezaurs.lv, Korpuss.lv, the LV-PIPE tools, CLARIN-LV and Balsutalka.lv. Its FullStack-LV repository offers a multilayer corpus with Universal Dependencies, FrameNet, PropBank and AMR layers, released under a Creative Commons licence.",
    "Funding comes from the European Union, the European Regional Development Fund and Latvian ministries and councils. The lab has partnered with the news agency LETA, Deutsche Welle and the National Library of Latvia."
   ],
   "mission": "The lab states that it develops resources and technologies that help Latvian thrive in a multilingual environment. It publishes its tools and data openly for researchers and developers.",
   "notice": null,
   "known_for": [
    "Tezaurs Latvian lexical database",
    "FullStack-LV multilayer corpus",
    "LVBERT Latvian language model",
    "SemEval 2016 and 2017 results"
   ],
   "focus": [
    "Latvian NLP",
    "lexicons",
    "language models"
   ],
   "contributions": [
    {
     "title": "Tezaurs repository",
     "note": "The open Latvian lexical database repository was created on GitHub in April 2015.",
     "year": "2015"
    },
    {
     "title": "SemEval 2016 semantic analysis",
     "note": "The lab reports the best result in the semantic analysis task at SemEval 2016.",
     "year": "2016"
    },
    {
     "title": "SemEval 2017 text generation",
     "note": "The lab reports the best result in the text generation task at SemEval 2017.",
     "year": "2017"
    },
    {
     "title": "FullStack-LV",
     "note": "A multilayer Latvian corpus with UD, FrameNet, PropBank and AMR layers was started and published at LREC 2018.",
     "year": "2018"
    },
    {
     "title": "LVBERT",
     "note": "A pretrained Latvian BERT model repository was created in July 2020.",
     "year": "2020"
    },
    {
     "title": "MultiCoNER 2023",
     "note": "The lab placed third among European teams in the named entity recognition task.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "Tezaurs",
     "url": "https://github.com/LUMII-AILab/Tezaurs",
     "note": "The largest open lexical database of Latvian (65 stars)"
    },
    {
     "name": "FullStack",
     "url": "https://github.com/LUMII-AILab/FullStack",
     "note": "Full Stack of Latvian Language Resources for Natural Language Understanding (NLU) and Generation (NLG) (16 stars)"
    },
    {
     "name": "LVBERT",
     "url": "https://github.com/LUMII-AILab/LVBERT",
     "note": "Pretrained contextualized transformer-based state-of-the-art language model for Latvian (14 stars)"
    },
    {
     "name": "NLP_Course",
     "url": "https://github.com/LUMII-AILab/NLP_Course",
     "note": "Studiju kursu \"Valodu tehnoloģiju pamati\" un \"Valodu tehnoloģiju lietojumi\" praktisko darbu materiāli (10 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Tezaurs",
     "url": "https://github.com/LUMII-AILab/Tezaurs",
     "note": "Largest open lexical database of Latvian"
    },
    {
     "name": "FullStack",
     "url": "https://github.com/LUMII-AILab/FullStack",
     "note": "Full stack of Latvian language resources"
    },
    {
     "name": "LVBERT",
     "url": "https://github.com/LUMII-AILab/LVBERT",
     "note": "Latvian BERT model"
    }
   ],
   "openness": "Language resources and models are public on GitHub.",
   "key_people": [
    {
     "name": "Normunds Grūzītis",
     "role": "Coordinator of the FullStack-LV project",
     "url": "http://www.linkedin.com/in/normundsg",
     "note": null
    },
    {
     "name": "Pēteris Paikens",
     "role": "Researcher and member of the LUMII-AILab GitHub organization",
     "url": "https://github.com/PeterisP",
     "note": null
    },
    {
     "name": "Lauma Pretkalniņa",
     "role": "Researcher, enhanced UD from the hybrid treebank",
     "url": "https://github.com/lauma",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the address printed on the lab website.",
   "sources": [
    "https://ailab.lv",
    "https://github.com/LUMII-AILab"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "ailab",
   "mlat": 56.91424,
   "mlon": 24.1052,
   "logo": true
  },
  {
   "orig_id": "aiml-tuda",
   "name": "AIML Lab",
   "alias": "TU Darmstadt",
   "full_name": "AIML Lab (TU Darmstadt)",
   "type": "academic",
   "founded": null,
   "country": "Germany",
   "city": "Darmstadt",
   "lat": 49.8728,
   "lon": 8.6512,
   "website": "https://www.aiml.informatik.tu-darmstadt.de",
   "huggingface": "https://huggingface.co/AIML-TUDA",
   "github": "https://github.com/ml-research",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Artificial Intelligence and Machine Learning Lab at TU Darmstadt, open-sourcing safety tooling for vision-language and diffusion models such as LlavaGuard and Safe Latent Diffusion.",
   "tagline": "The AI and Machine Learning Lab at TU Darmstadt releases open safety tooling and reasoning research, including LlavaGuard and Safe Latent Diffusion.",
   "about": [
    "The AIML Lab at TU Darmstadt is led by Professor Kristian Kersting, who joined the university in 2019 and also co-directs hessian.AI. Its stated aims are advancing statistical learning, probabilistic reasoning and collaborative intelligence, with work in statistical relational AI, generative modeling and trustworthy machine learning.",
    "The lab publishes code on GitHub and models on Hugging Face, where it lists 20 models and 42 datasets. Safety tools include Safe Latent Diffusion, LlavaGuard and QwenGuard, and recent releases cover OlmoLogic reasoning models and the KletterMix German pretraining data.",
    "Its audience is researchers and developers who need open checks for image-text datasets and generative models."
   ],
   "mission": "The lab aims to make machines learn about the world rapidly and flexibly, combining logical and probabilistic methods with deep learning. It emphasises trustworthy and safe generative systems.",
   "notice": null,
   "known_for": [
    "Safe Latent Diffusion",
    "LlavaGuard vision-language safeguards",
    "Statistical relational AI",
    "OlmoLogic reasoning models"
   ],
   "focus": [
    "AI safety",
    "vision-language"
   ],
   "contributions": [
    {
     "title": "Safe Latent Diffusion",
     "note": "Released a method and model for mitigating inappropriate image generation (arXiv 2211.05105).",
     "year": "2022"
    },
    {
     "title": "LlavaGuard",
     "note": "Released open vision-language safeguards for datasets and models (arXiv 2406.05113).",
     "year": "2024"
    },
    {
     "title": "QwenGuard",
     "note": "Published further image-text safety models alongside LlavaGuard on Hugging Face.",
     "year": "2025"
    },
    {
     "title": "OlmoLogic",
     "note": "Published reasoning models trained with RLVR and inductive logic programming.",
     "year": "2026"
    },
    {
     "title": "KletterMix",
     "note": "Released a German pretraining data collection.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "semantic-image-editing",
     "url": "https://github.com/ml-research/semantic-image-editing",
     "note": "Open-source repository. (209 stars)"
    },
    {
     "name": "liground",
     "url": "https://github.com/ml-research/liground",
     "note": "A free, open-source and modern Chess Variant Analysis GUI for the 21st century (160 stars)"
    },
    {
     "name": "rational_activations",
     "url": "https://github.com/ml-research/rational_activations",
     "note": "Rational Activation Functions - Replacing Padé Activation Units (107 stars)"
    },
    {
     "name": "safe-latent-diffusion",
     "url": "https://github.com/ml-research/safe-latent-diffusion",
     "note": "Official Implementation of Safe Latent Diffusion for Text2Image (97 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "AIML-TUDA/stable-diffusion-safe",
     "url": "https://huggingface.co/AIML-TUDA/stable-diffusion-safe",
     "kind": "model",
     "note": "Model with 15 likes on the Hub."
    },
    {
     "name": "AIML-TUDA/LlavaGuard-v1.0-7B",
     "url": "https://huggingface.co/AIML-TUDA/LlavaGuard-v1.0-7B",
     "kind": "model",
     "note": "Model with 11 likes on the Hub."
    },
    {
     "name": "AIML-TUDA/i2p",
     "url": "https://huggingface.co/datasets/AIML-TUDA/i2p",
     "kind": "dataset",
     "note": "Dataset with 21 likes on the Hub."
    },
    {
     "name": "AIML-TUDA/semantic-diffusion",
     "url": "https://huggingface.co/spaces/AIML-TUDA/semantic-diffusion",
     "kind": "space",
     "note": "Space with 29 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "LlavaGuard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models",
     "url": "https://arxiv.org/abs/2406.05113",
     "year": "2024"
    },
    {
     "title": "Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models",
     "url": "https://arxiv.org/abs/2211.05105",
     "year": "2022"
    },
    {
     "title": "SEGA: Instructing Text-to-Image Models using Semantic Guidance",
     "url": "https://arxiv.org/abs/2301.12247",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "LlavaGuard v1.2 7B",
     "url": "https://huggingface.co/AIML-TUDA/LlavaGuard-v1.2-7B-OV-hf",
     "note": "Vision-language safety model"
    },
    {
     "name": "stable-diffusion-safe",
     "url": "https://huggingface.co/AIML-TUDA/stable-diffusion-safe",
     "note": "Safe Latent Diffusion checkpoint"
    },
    {
     "name": "I2P",
     "url": "https://huggingface.co/datasets/AIML-TUDA/i2p",
     "note": "Inappropriate image prompts dataset"
    },
    {
     "name": "semantic-image-editing",
     "url": "https://github.com/ml-research/semantic-image-editing",
     "note": "Code for semantic guidance in diffusion"
    }
   ],
   "openness": "Models, datasets and code are public on Hugging Face and GitHub (ml-research).",
   "key_people": [
    {
     "name": "Kristian Kersting",
     "role": "Professor of AI and Machine Learning, head of the lab",
     "url": "https://ml-research.github.io/people/kkersting/",
     "note": "Also heads a research department at DFKI."
    },
    {
     "name": "Patrick Schramowski",
     "role": "Researcher; lead author on Safe Latent Diffusion and LlavaGuard",
     "url": "https://arxiv.org/abs/2406.05113",
     "note": null
    },
    {
     "name": "Manuel Brack",
     "role": "Researcher; co-author of Safe Latent Diffusion and LlavaGuard",
     "url": "https://arxiv.org/abs/2211.05105",
     "note": null
    },
    {
     "name": "Quentin Delfosse",
     "role": "Researcher at TU Darmstadt",
     "url": "https://www.ml.informatik.tu-darmstadt.de/people/qdelfosse/index.html",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.aiml.informatik.tu-darmstadt.de",
    "https://huggingface.co/AIML-TUDA",
    "https://github.com/ml-research",
    "https://arxiv.org/abs/2406.05113"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "aiml-lab",
   "mlat": 49.89445,
   "mlon": 8.70939,
   "logo": true
  },
  {
   "orig_id": "aleph-alpha",
   "name": "Aleph Alpha",
   "alias": null,
   "full_name": "Aleph Alpha",
   "type": "startup",
   "founded": 2019,
   "country": "Germany",
   "city": "Heidelberg",
   "lat": 49.3988,
   "lon": 8.6724,
   "website": "https://aleph-alpha.com",
   "huggingface": "https://huggingface.co/Aleph-Alpha",
   "github": "https://github.com/Aleph-Alpha",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Heidelberg language-model company that publishes part of its research as open weights and datasets alongside commercial products.",
   "tagline": "Heidelberg language-model company that built sovereign European models and released part of its research as open weights, now being merged into Cohere.",
   "about": [
    "Aleph Alpha was founded in 2019 in Heidelberg by Jonas Andrulis and Samuel Weinbach. It raised a 5.3 million euro seed round in 2020, a further 23 million euros in 2021, and in November 2023 announced a large round led by Schwarz Gruppe and the Dieter Schwarz Foundation.",
    "The company sells specialised sovereign language models and solutions to organisations, and also publishes research code, datasets and models. Its Hugging Face organization carries the Kolibri model, a German and English mixture-of-experts model released in October 2026 under Apache 2.0.",
    "In April 2026 Cohere announced that it would acquire Aleph Alpha, and a definitive merger agreement followed in September 2026. The combined company will operate as Cohere, with Heidelberg kept as a research centre, subject to regulatory approval."
   ],
   "mission": "The company says it builds specialised sovereign models that give customers control over their data and adapt to their context. Its open releases support German-language and tokenizer research.",
   "notice": "Cohere and Aleph Alpha signed a definitive combination agreement on 16 September 2026. The combined company is to operate as Cohere, with Aleph Alpha's Heidelberg office focused on research. The deal is expected to close later in 2026, subject to regulatory approval.",
   "known_for": [
    "Sovereign European LLMs",
    "T-FREE tokenizer-free models",
    "GermanWeb pre-training corpus",
    "Kolibri open-weight model"
   ],
   "focus": [
    "language models",
    "German",
    "tokenization"
   ],
   "contributions": [
    {
     "title": "Founding in Heidelberg",
     "note": "Jonas Andrulis and Samuel Weinbach founded the company.",
     "year": "2019"
    },
    {
     "title": "Seed funding",
     "note": "The company secured 5.3 million euros in seed funding.",
     "year": "2020"
    },
    {
     "title": "T-FREE",
     "note": "Published a tokenizer-free approach to language models with code (arXiv 2406.19223).",
     "year": "2024"
    },
    {
     "title": "GermanWeb",
     "note": "Released a German pre-training corpus and its paper (arXiv 2505.00022).",
     "year": "2025"
    },
    {
     "title": "Kolibri",
     "note": "Released a German and English mixture-of-experts model with Apache 2.0 weights.",
     "year": "2026"
    },
    {
     "title": "Cohere merger",
     "note": "A definitive agreement was signed to merge the company into Cohere.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "ts-rs",
     "url": "https://github.com/Aleph-Alpha/ts-rs",
     "note": "Generate TypeScript bindings from Rust types (1,958 stars)"
    },
    {
     "name": "aleph-alpha-client",
     "url": "https://github.com/Aleph-Alpha/aleph-alpha-client",
     "note": "Python Client for the Aleph Alpha API (105 stars)"
    },
    {
     "name": "intelligence-layer-sdk",
     "url": "https://github.com/Aleph-Alpha/intelligence-layer-sdk",
     "note": "a unified framework for leveraging LLMs (79 stars)"
    },
    {
     "name": "Alpha-MoE",
     "url": "https://github.com/Aleph-Alpha/Alpha-MoE",
     "note": "Open-source repository. (77 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "Aleph-Alpha/Kolibri-1",
     "url": "https://huggingface.co/Aleph-Alpha/Kolibri-1",
     "kind": "model",
     "note": "Model with 1,013 likes on the Hub."
    },
    {
     "name": "Aleph-Alpha/Pharia-1-LLM-7B-control",
     "url": "https://huggingface.co/Aleph-Alpha/Pharia-1-LLM-7B-control",
     "kind": "model",
     "note": "Model with 71 likes on the Hub."
    },
    {
     "name": "Aleph-Alpha/Aleph-Alpha-GermanWeb",
     "url": "https://huggingface.co/datasets/Aleph-Alpha/Aleph-Alpha-GermanWeb",
     "kind": "dataset",
     "note": "Dataset with 28 likes on the Hub."
    },
    {
     "name": "Aleph-Alpha/meta-eval-argilla",
     "url": "https://huggingface.co/spaces/Aleph-Alpha/meta-eval-argilla",
     "kind": "space",
     "note": "Space with 0 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "T-FREE: Subword Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings",
     "url": "https://arxiv.org/abs/2406.19223",
     "year": "2024"
    },
    {
     "title": "Aleph-Alpha-GermanWeb: Improving German-language LLM pre-training with model-based data curation and synthetic data generation",
     "url": "https://arxiv.org/abs/2505.00022",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Kolibri-1",
     "url": "https://huggingface.co/Aleph-Alpha/Kolibri-1",
     "note": "German-English model with an Apache-2.0 card"
    },
    {
     "name": "Pharia-1-LLM-7B-control",
     "url": "https://huggingface.co/Aleph-Alpha/Pharia-1-LLM-7B-control",
     "note": "Open-weight 7B model under a custom open licence"
    },
    {
     "name": "Aleph-Alpha-GermanWeb",
     "url": "https://huggingface.co/datasets/Aleph-Alpha/Aleph-Alpha-GermanWeb",
     "note": "German pre-training dataset"
    },
    {
     "name": "Alpha-MoE",
     "url": "https://github.com/Aleph-Alpha/Alpha-MoE",
     "note": "Mixture-of-experts kernels and code"
    }
   ],
   "openness": "Mixed: flagship products are commercial, but model weights, datasets and code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Jonas Andrulis",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Aleph_Alpha",
     "note": null
    },
    {
     "name": "Samuel Weinbach",
     "role": "Co-founder and Co-Chief Research Officer",
     "url": "https://aleph-alpha.com/about-us/",
     "note": null
    },
    {
     "name": "Ilhan Scheer",
     "role": "Chief Executive Officer",
     "url": "https://aleph-alpha.com/about-us/",
     "note": null
    },
    {
     "name": "Yasser Jadidi",
     "role": "Co-Chief Research Officer",
     "url": "https://aleph-alpha.com/about-us/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://aleph-alpha.com",
    "https://huggingface.co/Aleph-Alpha",
    "https://github.com/Aleph-Alpha",
    "https://arxiv.org/abs/2406.19223"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "aleph-alpha",
   "mlat": 49.36344,
   "mlon": 8.6724,
   "logo": true
  },
  {
   "orig_id": "alexandra-institute",
   "name": "Alexandra Institute",
   "alias": null,
   "full_name": "Alexandra Institute",
   "type": "nonprofit",
   "founded": 1999,
   "country": "Denmark",
   "city": "Aarhus",
   "lat": 56.1629,
   "lon": 10.2039,
   "website": "https://alexandra.dk",
   "huggingface": "https://huggingface.co/alexandrainst",
   "github": "https://github.com/alexandrainst",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Danish non-profit research and technology institute with open Danish NLP resources including DaNLP and the ScandEval benchmark.",
   "tagline": "Danish non-profit research and technology institute whose open resources include DaNLP and the ScandEval benchmark, now EuroEval.",
   "about": [
    "The Alexandra Institute was established in Aarhus in 1999 by business people, researchers and public representatives to link Danish companies with university IT research. It became an approved technological service provider in 2007 and has been owned by the Aarhus University Research Foundation since 2012.",
    "Its open source work is published on GitHub and Hugging Face. DaNLP collected Danish datasets and models for tagging, named entity recognition, sentiment and hate speech detection, with a NoDaLiDa 2021 paper; the repository was archived in March 2025. The ScandEval benchmark, described in arXiv 2304.00906, later grew into EuroEval.",
    "The institute serves Danish companies and public bodies and works on language technology, evaluation and responsible AI."
   ],
   "mission": "The institute applies IT research to the needs of Danish business and society and publishes language resources so that Danish remains well supported in AI.",
   "notice": null,
   "known_for": [
    "DaNLP Danish NLP resources",
    "ScandEval and EuroEval benchmark",
    "Open Danish datasets and models",
    "Responsible AI research"
   ],
   "focus": [
    "Danish NLP",
    "evaluation",
    "responsible AI"
   ],
   "contributions": [
    {
     "title": "Institute founded",
     "note": "The Alexandra Institute was established in Aarhus to connect business and IT research.",
     "year": "1999"
    },
    {
     "title": "GTS approval",
     "note": "It was approved as a technological service provider in Denmark.",
     "year": "2007"
    },
    {
     "title": "DaNLP repository",
     "note": "The Danish NLP resource repository was created on GitHub.",
     "year": "2019"
    },
    {
     "title": "DaNLP paper",
     "note": "The DaNLP paper by Pauli, Barrett, Lacroix and Hvingelby was presented at NoDaLiDa.",
     "year": "2021"
    },
    {
     "title": "ScandEval",
     "note": "The Scandinavian language model benchmark was described in arXiv 2304.00906.",
     "year": "2023"
    },
    {
     "title": "EuroEval",
     "note": "The benchmark continued under the EuroEval name, with a repository created in September 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "danlp",
     "url": "https://github.com/alexandrainst/danlp",
     "note": "DaNLP is a repository for Natural Language Processing resources for the Danish Language. (209 stars)"
    },
    {
     "name": "responsible-ai",
     "url": "https://github.com/alexandrainst/responsible-ai",
     "note": "Responsible AI knowledge base (119 stars)"
    },
    {
     "name": "processing_websockets",
     "url": "https://github.com/alexandrainst/processing_websockets",
     "note": "A web socket library, including both server and client, for Processing (95 stars)"
    },
    {
     "name": "alexandra-trackmap-panel",
     "url": "https://github.com/alexandrainst/alexandra-trackmap-panel",
     "note": "Grafana map plugin to visualise coordinates as markers, hexbin, ant path, or heatmap. (79 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "alexandrainst/scandi-nli-large",
     "url": "https://huggingface.co/alexandrainst/scandi-nli-large",
     "kind": "model",
     "note": "Model with 11 likes on the Hub."
    },
    {
     "name": "alexandrainst/da-sentiment-base",
     "url": "https://huggingface.co/alexandrainst/da-sentiment-base",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "alexandrainst/multi-wiki-qa",
     "url": "https://huggingface.co/datasets/alexandrainst/multi-wiki-qa",
     "kind": "dataset",
     "note": "Dataset with 25 likes on the Hub."
    },
    {
     "name": "alexandrainst/radial-plot-generator",
     "url": "https://huggingface.co/spaces/alexandrainst/radial-plot-generator",
     "kind": "space",
     "note": "Space with 7 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "ScandEval: A Benchmark for Scandinavian Natural Language Processing",
     "url": "https://arxiv.org/abs/2304.00906",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "DaNLP",
     "url": "https://github.com/alexandrainst/danlp",
     "note": "Resources for Danish NLP"
    },
    {
     "name": "scandi-nli",
     "url": "https://huggingface.co/alexandrainst/scandi-nli-large",
     "note": "Scandinavian NLI model"
    },
    {
     "name": "multi-wiki-qa",
     "url": "https://huggingface.co/datasets/alexandrainst/multi-wiki-qa",
     "note": "Multilingual QA dataset"
    },
    {
     "name": "responsible-ai",
     "url": "https://github.com/alexandrainst/responsible-ai",
     "note": "Responsible AI knowledge base"
    }
   ],
   "openness": "Models, datasets and tools are public under the alexandrainst accounts on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Niels Husted Kjær",
     "role": "Chief Executive Officer per the GTS institute listing",
     "url": "https://gts-net.dk/english/the-gts-institutes/alexandra-instituttet-en/",
     "note": null
    },
    {
     "name": "Peter Andersen",
     "role": "Member of the alexandrainst GitHub organization",
     "url": "https://github.com/datpete",
     "note": null
    },
    {
     "name": "Jacob Harder",
     "role": "Member of the alexandrainst GitHub organization",
     "url": "https://github.com/harderj",
     "note": null
    },
    {
     "name": "Alexandre Alapetite",
     "role": "Member of the alexandrainst GitHub organization",
     "url": "https://alexandre.alapetite.fr/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the institute website.",
   "sources": [
    "https://alexandra.dk",
    "https://huggingface.co/alexandrainst",
    "https://github.com/alexandrainst",
    "https://arxiv.org/abs/2304.00906"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "alexandra-institute",
   "mlat": 56.1196,
   "mlon": 10.2039,
   "logo": true
  },
  {
   "orig_id": "allen-institute-for-ai",
   "name": "Allen Institute for AI",
   "alias": "Ai2",
   "full_name": "Allen Institute for AI (Ai2)",
   "type": "nonprofit",
   "founded": 2014,
   "country": "United States",
   "city": "Seattle",
   "lat": 47.6062,
   "lon": -122.3321,
   "website": "https://allenai.org",
   "huggingface": "https://huggingface.co/allenai",
   "github": "https://github.com/allenai",
   "x": "https://x.com/allen_ai",
   "discord": "https://discord.gg/ai2",
   "linkedin": null,
   "contact_email": "info@allenai.org",
   "summary": "Seattle-based nonprofit research institute building fully open language models (OLMo), datasets (Dolma), post-training recipes (Tulu) and science and vision tools.",
   "tagline": "A Seattle nonprofit research institute founded by Paul Allen that builds fully open language models, datasets and training recipes.",
   "about": [
    "The Allen Institute for AI, known as Ai2, was founded in 2014 by Microsoft co-founder Paul Allen. Oren Etzioni led it for nine years until 2022, and Ali Farhadi became CEO in July 2023.",
    "Its early work included the Semantic Scholar search engine, released in November 2015, and the Aristo project, which passed an eighth-grade science exam in 2018. Since 2023 it has concentrated on open language models.",
    "The OLMo models and the Dolma corpus were released in February 2024 with data, code and checkpoints, and the Tulu post-training recipes reached version 3 in November 2024."
   ],
   "mission": "Ai2 pursues AI research and engineering for the common good, with open release of the models, data and code behind its work.",
   "notice": null,
   "known_for": [
    "OLMo fully open language models",
    "Dolma open pretraining corpus",
    "Tulu open post-training recipes",
    "Semantic Scholar"
   ],
   "focus": [
    "language models",
    "datasets",
    "post-training",
    "science"
   ],
   "contributions": [
    {
     "title": "Semantic Scholar released",
     "note": "An AI-backed academic search engine.",
     "year": "2015"
    },
    {
     "title": "Aristo passes science exam",
     "note": "Its system achieved a pass on an eighth-grade science exam.",
     "year": "2018"
    },
    {
     "title": "OLMo and Dolma released",
     "note": "1B and 7B models with a three-trillion-token open corpus.",
     "year": "2024"
    },
    {
     "title": "Tulu 3 released",
     "note": "An open post-training recipe following Tulu in 2023 and Tulu 2 later that year.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "olmocr",
     "url": "https://github.com/allenai/olmocr",
     "note": "Most-starred repo; toolkit for linearizing PDFs for LLM datasets."
    },
    {
     "name": "allennlp",
     "url": "https://github.com/allenai/allennlp",
     "note": "Long-running research NLP library on PyTorch."
    },
    {
     "name": "OLMo",
     "url": "https://github.com/allenai/OLMo",
     "note": "Modeling, training and evaluation code for OLMo."
    },
    {
     "name": "open-instruct",
     "url": "https://github.com/allenai/open-instruct",
     "note": "Post-training codebase behind Tulu."
    }
   ],
   "hf_highlights": [
    {
     "name": "longformer-base-4096",
     "url": "https://huggingface.co/allenai/longformer-base-4096",
     "kind": "model",
     "note": "Longformer for long documents."
    },
    {
     "name": "specter2_base",
     "url": "https://huggingface.co/allenai/specter2_base",
     "kind": "model",
     "note": "Scientific document embedding model."
    },
    {
     "name": "c4",
     "url": "https://huggingface.co/datasets/allenai/c4",
     "kind": "dataset",
     "note": "Cleaned C4 corpus mirror."
    },
    {
     "name": "ai2_arc",
     "url": "https://huggingface.co/datasets/allenai/ai2_arc",
     "kind": "dataset",
     "note": "ARC science question benchmark."
    }
   ],
   "papers": [
    {
     "title": "OLMo: Accelerating the Science of Language Models",
     "url": "https://arxiv.org/abs/2402.00838",
     "year": "2024"
    },
    {
     "title": "2 OLMo 2 Furious",
     "url": "https://arxiv.org/abs/2501.00656",
     "year": "2024"
    },
    {
     "title": "Tulu 3: Pushing Frontiers in Open Language Model Post-Training",
     "url": "https://arxiv.org/abs/2411.15124",
     "year": "2024"
    },
    {
     "title": "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research",
     "url": "https://arxiv.org/abs/2402.00159",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "OLMo",
     "url": "https://github.com/allenai/OLMo",
     "note": "Fully open LLM code, data and checkpoints"
    },
    {
     "name": "open-instruct",
     "url": "https://github.com/allenai/open-instruct",
     "note": "Post-training codebase"
    },
    {
     "name": "olmocr",
     "url": "https://github.com/allenai/olmocr",
     "note": "PDF to text toolkit"
    }
   ],
   "openness": "Releases weights, training code and data, mostly under Apache-2.0.",
   "key_people": [
    {
     "name": "Ali Farhadi",
     "role": "CEO from 2023 (Wikipedia notes he stepped down in March 2026)",
     "url": "https://allenai.org/",
     "note": null
    },
    {
     "name": "Peter Clark",
     "role": "Senior researcher and interim CEO",
     "url": "https://allenai.org/",
     "note": null
    },
    {
     "name": "Oren Etzioni",
     "role": "Former CEO",
     "url": "https://en.wikipedia.org/wiki/Oren_Etzioni",
     "note": null
    },
    {
     "name": "Hannaneh Hajishirzi",
     "role": "Researcher",
     "url": "https://allenai.org/",
     "note": null
    },
    {
     "name": "Yejin Choi",
     "role": "Researcher",
     "url": "https://en.wikipedia.org/wiki/Yejin_Choi",
     "note": null
    },
    {
     "name": "Noah Smith",
     "role": "Researcher",
     "url": "https://allenai.org/",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://allenai.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://allenai.org",
    "https://huggingface.co/allenai",
    "https://github.com/allenai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "ai2",
   "mlat": 47.57084,
   "mlon": -122.3321,
   "logo": true
  },
  {
   "orig_id": "almanach",
   "name": "ALMAnaCH",
   "alias": "Inria",
   "full_name": "ALMAnaCH (Inria)",
   "type": "institute",
   "founded": null,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://almanach.inria.fr",
   "huggingface": "https://huggingface.co/almanach",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Inria Paris research team on language models and computational humanities, creators of the CamemBERT family of French language models.",
   "tagline": "Inria Paris research team behind the CamemBERT family of open French language models and work in computational humanities.",
   "about": [
    "ALMAnaCH is an Inria team in Paris working on natural language processing and digital humanities. Its members include Benoît Sagot and Djamé Seddah, who are authors of the original CamemBERT paper alongside colleagues from Inria and Sorbonne Université.",
    "CamemBERT, a RoBERTa-based French model released under an MIT licence in 2019, became a widely adopted reference. CamemBERT 2.0 followed in 2024, and CamemBERTa, by Wissam Antoun, Sagot and Seddah, explored data-efficient training. The team's Hugging Face organization lists 92 models and 43 datasets, including biomedical encoders and Gaperon sparse autoencoders.",
    "The models serve researchers, companies and public bodies that need open French and biomedical text processing."
   ],
   "mission": "The team builds and shares language models and data for French and historical or specialised text. It treats open release as part of academic practice.",
   "notice": null,
   "known_for": [
    "CamemBERT French language model",
    "CamemBERT 2.0 and CamemBERTa",
    "French biomedical encoders",
    "Digital humanities NLP"
   ],
   "focus": [
    "NLP",
    "French",
    "language models",
    "digital humanities"
   ],
   "contributions": [
    {
     "title": "CamemBERT",
     "note": "Martin et al. released the reference open French language model (arXiv 1911.03894).",
     "year": "2019"
    },
    {
     "title": "CamemBERTa",
     "note": "Antoun, Sagot and Seddah published a data-efficient French language model.",
     "year": "2023"
    },
    {
     "title": "CamemBERT 2.0",
     "note": "Updated French encoders with a refreshed tokenizer and data were released (arXiv 2411.08868).",
     "year": "2024"
    },
    {
     "title": "Biomedical models and data",
     "note": "Published French biomedical encoders and datasets such as Biomed-Enriched.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "almanach/camembert-base",
     "url": "https://huggingface.co/almanach/camembert-base",
     "kind": "model",
     "note": "Model with 117 likes on the Hub."
    },
    {
     "name": "almanach/camembertav2-base",
     "url": "https://huggingface.co/almanach/camembertav2-base",
     "kind": "model",
     "note": "Model with 28 likes on the Hub."
    },
    {
     "name": "almanach/Biomed-Enriched",
     "url": "https://huggingface.co/datasets/almanach/Biomed-Enriched",
     "kind": "dataset",
     "note": "Dataset with 10 likes on the Hub."
    },
    {
     "name": "almanach/camembert-bio-gliner",
     "url": "https://huggingface.co/spaces/almanach/camembert-bio-gliner",
     "kind": "space",
     "note": "Space with 10 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "CamemBERT: a Tasty French Language Model",
     "url": "https://arxiv.org/abs/1911.03894",
     "year": "2019"
    },
    {
     "title": "CamemBERT 2.0: A Smarter French Language Model Aged to Perfection",
     "url": "https://arxiv.org/abs/2411.08868",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "CamemBERT",
     "url": "https://huggingface.co/almanach/camembert-base",
     "note": "Widely used French RoBERTa-style model"
    },
    {
     "name": "CamemBERT-bio",
     "url": "https://huggingface.co/almanach/camembert-bio-base",
     "note": "French biomedical encoder"
    },
    {
     "name": "Biomed-Enriched",
     "url": "https://huggingface.co/datasets/almanach/Biomed-Enriched",
     "note": "Biomedical dataset released on the Hub"
    },
    {
     "name": "CamemBERT 2.0",
     "url": "https://huggingface.co/almanach/camembertav2-base",
     "note": "Updated French encoder"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face under the almanach organisation; CamemBERT checkpoints have very high download counts.",
   "key_people": [
    {
     "name": "Benoît Sagot",
     "role": "Researcher, CamemBERT author",
     "url": "https://huggingface.co/almanach",
     "note": null
    },
    {
     "name": "Djamé Seddah",
     "role": "Senior researcher at Inria",
     "url": "https://www.alphaxiv.org/@djame-seddah",
     "note": null
    },
    {
     "name": "Wissam Antoun",
     "role": "Researcher, CamemBERTa author",
     "url": "https://github.com/WissamAntoun",
     "note": null
    },
    {
     "name": "Thibault Clérice",
     "role": "Member of the almanach Hugging Face organization",
     "url": "https://huggingface.co/almanach",
     "note": null
    },
    {
     "name": "Nathan Godey",
     "role": "Member of the almanach Hugging Face organization",
     "url": "https://huggingface.co/almanach",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the Inria team page.",
   "sources": [
    "https://almanach.inria.fr",
    "https://huggingface.co/almanach",
    "https://arxiv.org/abs/1911.03894"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "almanach",
   "mlat": 48.75352,
   "mlon": 2.3522,
   "logo": true
  },
  {
   "orig_id": "answer-ai",
   "name": "Answer.AI",
   "alias": null,
   "full_name": "Answer.AI",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.answer.ai",
   "huggingface": "https://huggingface.co/answerdotai",
   "github": "https://github.com/AnswerDotAI",
   "x": "https://x.com/answerdotai",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Practical AI R&D lab founded by Jeremy Howard and Eric Ries as a public benefit corporation, shipping open-source libraries and models such as ModernBERT, FastHTML, nbdev and RAGatouille.",
   "tagline": "A public benefit AI R&D lab that turns research advances into open-source libraries, models and tools for developers.",
   "about": [
    "Answer.AI announced itself in December 2023 in a post by Jeremy Howard describing a new kind of R&D lab that builds practical products from foundational research. In November 2024 fast.ai announced that it was joining the lab.",
    "Its output is mostly open source. Projects include a method for training 70B models on consumer GPUs with FSDP and QLoRA, the FastHTML web framework, the llms.txt proposal, and ModernBERT, an encoder model released in December 2024.",
    "The lab's team includes Jeremy Howard, Hamel Husain, Benjamin Clavié, Johno Whitaker and Benjamin Warner among others."
   ],
   "mission": "Answer.AI describes its aim as creating practical end-user products based on foundational research breakthroughs.",
   "notice": null,
   "known_for": [
    "ModernBERT encoder model",
    "FSDP+QLoRA for 70B models on consumer GPUs",
    "FastHTML and fast.ai tooling",
    "llms.txt proposal"
   ],
   "focus": [
    "encoder models",
    "developer tools",
    "retrieval",
    "efficient fine-tuning"
   ],
   "contributions": [
    {
     "title": "Lab announced",
     "note": "Jeremy Howard published the founding post for the lab.",
     "year": "2023"
    },
    {
     "title": "FSDP+QLoRA released",
     "note": "A method for training 70B models on consumer GPUs.",
     "year": "2024"
    },
    {
     "title": "FastHTML released",
     "note": "A Python web framework from the lab.",
     "year": "2024"
    },
    {
     "title": "llms.txt proposed",
     "note": "A proposal for websites to publish LLM-readable summaries.",
     "year": "2024"
    },
    {
     "title": "ModernBERT released",
     "note": "A modern encoder model.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "fasthtml",
     "url": "https://github.com/AnswerDotAI/fasthtml",
     "note": "Python library for building HTML apps"
    },
    {
     "name": "nbdev",
     "url": "https://github.com/AnswerDotAI/nbdev",
     "note": "Notebook-driven software development"
    },
    {
     "name": "RAGatouille",
     "url": "https://github.com/AnswerDotAI/RAGatouille",
     "note": "Late-interaction retrieval for RAG"
    },
    {
     "name": "ModernBERT",
     "url": "https://github.com/AnswerDotAI/ModernBERT",
     "note": "Training code for ModernBERT"
    }
   ],
   "hf_highlights": [
    {
     "name": "ModernBERT-base",
     "url": "https://huggingface.co/answerdotai/ModernBERT-base",
     "kind": "model",
     "note": "Base encoder model"
    },
    {
     "name": "ModernBERT-large",
     "url": "https://huggingface.co/answerdotai/ModernBERT-large",
     "kind": "model",
     "note": "Large encoder model"
    },
    {
     "name": "answerai-colbert-small-v1",
     "url": "https://huggingface.co/answerdotai/answerai-colbert-small-v1",
     "kind": "model",
     "note": "Small ColBERT retrieval model"
    }
   ],
   "papers": [
    {
     "title": "Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference",
     "url": "https://arxiv.org/abs/2412.13663",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "fasthtml",
     "url": "https://github.com/AnswerDotAI/fasthtml",
     "note": "Python library for building HTML apps"
    },
    {
     "name": "nbdev",
     "url": "https://github.com/AnswerDotAI/nbdev",
     "note": "Notebook-driven software development"
    },
    {
     "name": "RAGatouille",
     "url": "https://github.com/AnswerDotAI/RAGatouille",
     "note": "Late-interaction retrieval for RAG"
    }
   ],
   "openness": "Apache-2.0 code on GitHub and Apache-2.0 model weights on Hugging Face.",
   "key_people": [
    {
     "name": "Jeremy Howard",
     "role": "Founder and CEO",
     "url": "https://en.wikipedia.org/wiki/Jeremy_Howard_(entrepreneur)",
     "note": null
    },
    {
     "name": "Benjamin Clavié",
     "role": "Researcher",
     "url": "https://www.answer.ai/",
     "note": null
    },
    {
     "name": "Johno Whitaker",
     "role": "Researcher",
     "url": "https://www.answer.ai/",
     "note": null
    },
    {
     "name": "Hamel Husain",
     "role": "Team member",
     "url": "https://www.answer.ai/",
     "note": null
    },
    {
     "name": "Benjamin Warner",
     "role": "Researcher",
     "url": "https://www.answer.ai/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.answer.ai",
    "https://www.answer.ai/overview.html",
    "https://github.com/AnswerDotAI",
    "https://huggingface.co/answerdotai"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "answer-ai",
   "mlat": 37.6549,
   "mlon": -122.4194,
   "logo": true
  },
  {
   "orig_id": "apollo-research",
   "name": "Apollo Research",
   "alias": null,
   "full_name": "Apollo Research",
   "type": "institute",
   "founded": 2023,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://www.apolloresearch.ai",
   "huggingface": "https://huggingface.co/apollo-research",
   "github": "https://github.com/apolloresearch",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "London-based AI safety research organisation studying deceptive behaviour, evaluations and mechanistic interpretability, publishing code and datasets on GitHub and Hugging Face.",
   "tagline": "London-based AI safety organisation that evaluates frontier models for scheming and publishes open evaluation and interpretability work.",
   "about": [
    "Apollo Research was founded in 2023 and is led by CEO and co-founder Marius Hobbhahn. Its stated goal is to secure frontier AI systems across development, deployment and governance, with a focus on scheming, where advanced systems covertly pursue misaligned objectives.",
    "In 2023 it presented at the UK AI Safety Summit and advised the Frontier AI Taskforce. In 2024 it tested OpenAI's o1 before release and published early evidence of in-context scheming in frontier models; in 2025 it ran evaluations for major labs and worked with OpenAI on anti-scheming interventions.",
    "In 2026 the organisation became a Public Benefit Corporation, opened a San Francisco office and began building Watcher, an agent monitoring tool. It publishes code on GitHub, including end-to-end sparse dictionary learning."
   ],
   "mission": "Apollo aims to reduce the risk of deceptive and misaligned AI through evaluations, interpretability research and governance advice. It works with developers and governments.",
   "notice": null,
   "known_for": [
    "Scheming evaluations of frontier models",
    "End-to-end sparse autoencoders",
    "Pre-deployment testing with OpenAI",
    "Loss of control playbook"
   ],
   "focus": [
    "AI safety",
    "evaluations",
    "interpretability"
   ],
   "contributions": [
    {
     "title": "Presented at AI Safety Summit",
     "note": "Apollo identified scheming as a serious risk and presented at the UK AI Safety Summit.",
     "year": "2023"
    },
    {
     "title": "End-to-end SAEs",
     "note": "Released training code and weights for end-to-end sparse dictionary learning (arXiv 2405.12241).",
     "year": "2024"
    },
    {
     "title": "o1 pre-deployment testing",
     "note": "Partnered with OpenAI to test o1 before public release.",
     "year": "2024"
    },
    {
     "title": "In-context scheming evaluations",
     "note": "Published evaluations showing frontier models can scheme in context (arXiv 2412.04984).",
     "year": "2024"
    },
    {
     "title": "Anti-scheming work",
     "note": "Partnered with OpenAI on anti-scheming interventions and wrote a loss of control playbook for governments.",
     "year": "2025"
    },
    {
     "title": "Public Benefit Corporation",
     "note": "Became a PBC, opened a San Francisco office and began building Watcher.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "insider-trading",
     "url": "https://github.com/ApolloResearch/insider-trading",
     "note": "Open-source repository. (66 stars)"
    },
    {
     "name": "deception-detection",
     "url": "https://github.com/ApolloResearch/deception-detection",
     "note": "Open-source repository. (61 stars)"
    },
    {
     "name": "e2e_sae",
     "url": "https://github.com/ApolloResearch/e2e_sae",
     "note": "Sparse Autoencoder Training Library (59 stars)"
    },
    {
     "name": "apd",
     "url": "https://github.com/ApolloResearch/apd",
     "note": "Attribution-based Parameter Decomposition (35 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "apollo-research/gpt2_noLN",
     "url": "https://huggingface.co/apollo-research/gpt2_noLN",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "apollo-research/e2e-saes-gpt2",
     "url": "https://huggingface.co/apollo-research/e2e-saes-gpt2",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "apollo-research/contrastive-belief-updates",
     "url": "https://huggingface.co/datasets/apollo-research/contrastive-belief-updates",
     "kind": "dataset",
     "note": "Dataset with 5 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Frontier Models are Capable of In-context Scheming",
     "url": "https://arxiv.org/abs/2412.04984",
     "year": "2024"
    },
    {
     "title": "Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning",
     "url": "https://arxiv.org/abs/2405.12241",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "insider-trading",
     "url": "https://github.com/apolloresearch/insider-trading",
     "note": "Deception evaluation environment"
    },
    {
     "name": "deception-detection",
     "url": "https://github.com/apolloresearch/deception-detection",
     "note": "Probes for detecting model deception"
    },
    {
     "name": "e2e_sae",
     "url": "https://github.com/apolloresearch/e2e_sae",
     "note": "Sparse autoencoder training library"
    },
    {
     "name": "APD",
     "url": "https://github.com/apolloresearch/apd",
     "note": "Attribution-based parameter decomposition"
    }
   ],
   "openness": "Research code, SAE weights and datasets are public; its evaluations of frontier models are published as papers.",
   "key_people": [
    {
     "name": "Marius Hobbhahn",
     "role": "CEO and co-founder",
     "url": "https://www.linkedin.com/in/marius-hobbhahn-128927175/",
     "note": null
    },
    {
     "name": "Chris Akin",
     "role": "Chief Operating Officer",
     "url": "https://www.linkedin.com/in/akinchristopher/",
     "note": null
    },
    {
     "name": "Alexander Meinke",
     "role": "Head of Research",
     "url": "https://www.linkedin.com/in/alexmeinke/",
     "note": null
    },
    {
     "name": "Jérémy Scheurer",
     "role": "Member of Technical Staff",
     "url": "https://www.linkedin.com/in/j%C3%A9r%C3%A9my-scheurer-927563b0/",
     "note": null
    },
    {
     "name": "Charlotte Stix",
     "role": "Chief Governance Officer",
     "url": "https://www.apolloresearch.ai/team",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website.",
   "sources": [
    "https://www.apolloresearch.ai",
    "https://huggingface.co/apollo-research",
    "https://github.com/apolloresearch",
    "https://arxiv.org/abs/2412.04984"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "apollo-research",
   "mlat": 51.44126,
   "mlon": -0.1278,
   "logo": true
  },
  {
   "orig_id": "arc-institute",
   "name": "Arc Institute",
   "alias": null,
   "full_name": "Arc Institute",
   "type": "nonprofit",
   "founded": 2021,
   "country": "United States",
   "city": "Palo Alto",
   "lat": 37.4419,
   "lon": -122.143,
   "website": "https://arcinstitute.org",
   "huggingface": "https://huggingface.co/arcinstitute",
   "github": "https://github.com/ArcInstitute",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "hello@arcinstitute.org",
   "summary": "Nonprofit biomedical research institute that releases open AI models for biology such as Evo 2, State and Stack.",
   "tagline": "A nonprofit biomedical research institute that builds open AI models for biology, including Evo 2 and virtual cell models.",
   "about": [
    "Arc Institute was founded in 2021 by Silvana Konermann, Patrick Hsu and Patrick Collison with initial funding commitments of 650 million dollars. It gives researchers long-term support without grant applications.",
    "In 2024 researchers led by Hsu and Brian Hie released Evo, a language model trained on biological sequence data, published in Science in November 2024. In April 2025 Arc announced the Virtual Cell Atlas with 10x Genomics and Ultima Genomics.",
    "The institute publishes models and datasets on Hugging Face and GitHub."
   ],
   "mission": "Arc aims to give scientists freedom to pursue long-horizon work on human disease, and increasingly applies machine learning to build virtual cell models.",
   "notice": null,
   "known_for": [
    "Evo genomic language models",
    "Virtual Cell Atlas",
    "Core investigator model for long-term research"
   ],
   "focus": [
    "biology",
    "foundation models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Institute founded",
     "note": "Konermann, Hsu and Collison started Arc.",
     "year": "2021"
    },
    {
     "title": "Evo published",
     "note": "A model trained on DNA, RNA and protein sequence, published in Science in November.",
     "year": "2024"
    },
    {
     "title": "Virtual Cell Atlas announced",
     "note": "A single-cell data atlas for virtual cell models, with 10x Genomics and Ultima Genomics.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "evo2",
     "url": "https://github.com/ArcInstitute/evo2",
     "note": "Genome modeling and design."
    },
    {
     "name": "state",
     "url": "https://github.com/ArcInstitute/state",
     "note": "Cell perturbation model."
    },
    {
     "name": "arc-virtual-cell-atlas",
     "url": "https://github.com/ArcInstitute/arc-virtual-cell-atlas",
     "note": "Virtual cell atlas."
    },
    {
     "name": "SRAgent",
     "url": "https://github.com/ArcInstitute/SRAgent",
     "note": "SRA agents."
    }
   ],
   "hf_highlights": [
    {
     "name": "evo2_7b",
     "url": "https://huggingface.co/arcinstitute/evo2_7b",
     "kind": "model",
     "note": "Evo 2 7B."
    },
    {
     "name": "evo2_40b",
     "url": "https://huggingface.co/arcinstitute/evo2_40b",
     "kind": "model",
     "note": "Evo 2 40B."
    },
    {
     "name": "opengenome2",
     "url": "https://huggingface.co/datasets/arcinstitute/opengenome2",
     "kind": "dataset",
     "note": "OpenGenome2 data."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Evo 2",
     "url": "https://github.com/ArcInstitute/evo2",
     "note": "Genome model"
    },
    {
     "name": "State",
     "url": "https://github.com/ArcInstitute/state",
     "note": "Perturbation model"
    }
   ],
   "openness": "Open weights and data.",
   "key_people": [
    {
     "name": "Silvana Konermann",
     "role": "Co-founder and executive director",
     "url": "https://arcinstitute.org/",
     "note": null
    },
    {
     "name": "Patrick Hsu",
     "role": "Co-founder and core investigator",
     "url": "https://arcinstitute.org/",
     "note": null
    },
    {
     "name": "Patrick Collison",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Patrick_Collison",
     "note": null
    },
    {
     "name": "Brian Hie",
     "role": "Evo lead",
     "url": "https://arcinstitute.org/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://arcinstitute.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://arcinstitute.org",
    "https://huggingface.co/arcinstitute"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "arc-institute",
   "logo": true
  },
  {
   "orig_id": "arc-prize",
   "name": "ARC Prize Foundation",
   "alias": null,
   "full_name": "ARC Prize Foundation",
   "type": "nonprofit",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://arcprize.org",
   "huggingface": "https://huggingface.co/arcprize",
   "github": "https://github.com/arcprize",
   "x": null,
   "discord": "https://discord.gg/9b77dPAmcA",
   "linkedin": null,
   "contact_email": null,
   "summary": "Nonprofit running the ARC-AGI benchmarks and prize competitions with open datasets and tooling.",
   "tagline": "A nonprofit that runs the ARC-AGI benchmarks and prize competitions to measure and encourage general reasoning in AI systems.",
   "about": [
    "ARC Prize Foundation is a nonprofit dedicated to accelerating progress toward AGI. It is built around ARC-AGI, the benchmark created by Francois Chollet, and runs open prize competitions on Kaggle.",
    "The foundation is led by co-founders Francois Chollet and Mike Knoop and president Greg Kamradt, who sit on the board with a small team working on benchmark design, engineering and model testing.",
    "Its site lists the ARC Prize 2024 and 2025 competitions and the ARC-AGI-2 and ARC-AGI-3 benchmarks, with ARC Prize 2026 results due on December 4, 2026."
   ],
   "mission": "The foundation states that it exists to guide researchers toward more general intelligence by publishing measurable, open benchmarks and prizes.",
   "notice": null,
   "known_for": [
    "ARC-AGI benchmarks",
    "ARC Prize competitions",
    "Open datasets for abstraction and reasoning"
   ],
   "focus": [
    "evaluation",
    "reasoning"
   ],
   "contributions": [
    {
     "title": "ARC Prize launched",
     "note": "The foundation began running open prize competitions on ARC-AGI.",
     "year": "2024"
    },
    {
     "title": "ARC-AGI-2 benchmark",
     "note": "A second benchmark release with a Kaggle competition.",
     "year": "2025"
    },
    {
     "title": "ARC-AGI-3 benchmark",
     "note": "A third benchmark listed on the foundation site.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "ARC-AGI-2",
     "url": "https://github.com/arcprize/ARC-AGI-2",
     "note": "ARC-AGI-2 tasks."
    },
    {
     "name": "arc-agi-benchmarking",
     "url": "https://github.com/arcprize/arc-agi-benchmarking",
     "note": "Baseline LLM testing."
    },
    {
     "name": "ARC-AGI-3-Agents",
     "url": "https://github.com/arcprize/ARC-AGI-3-Agents",
     "note": "ARC-AGI-3 agents."
    },
    {
     "name": "ARC-AGI",
     "url": "https://github.com/arcprize/ARC-AGI",
     "note": "ARC-AGI toolkit."
    }
   ],
   "hf_highlights": [
    {
     "name": "arc_agi_v2_public_eval",
     "url": "https://huggingface.co/datasets/arcprize/arc_agi_v2_public_eval",
     "kind": "dataset",
     "note": "ARC-AGI-2 public eval."
    },
    {
     "name": "arc_agi_v1_public_eval",
     "url": "https://huggingface.co/datasets/arcprize/arc_agi_v1_public_eval",
     "kind": "dataset",
     "note": "ARC-AGI-1 public eval."
    }
   ],
   "papers": [
    {
     "title": "ARC Prize 2024: Technical Report",
     "url": "https://arxiv.org/abs/2412.04604",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "ARC-AGI-2",
     "url": "https://github.com/arcprize/ARC-AGI-2",
     "note": "Benchmark"
    }
   ],
   "openness": "Open benchmark data and code.",
   "key_people": [
    {
     "name": "Francois Chollet",
     "role": "Co-founder, board member, creator of ARC-AGI",
     "url": "https://en.wikipedia.org/wiki/Fran%C3%A7ois_Chollet",
     "note": null
    },
    {
     "name": "Mike Knoop",
     "role": "Co-founder and board member",
     "url": "https://arcprize.org/about",
     "note": null
    },
    {
     "name": "Greg Kamradt",
     "role": "President and board member",
     "url": "https://arcprize.org/about",
     "note": null
    },
    {
     "name": "Bryan Landers",
     "role": "Founding team, design and strategy",
     "url": "https://arcprize.org/about",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://arcprize.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://arcprize.org",
    "https://huggingface.co/arcprize"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "arc-prize-foundation",
   "mlat": 37.65771,
   "mlon": -122.38676,
   "logo": true
  },
  {
   "orig_id": "arcee-ai",
   "name": "Arcee AI",
   "alias": null,
   "full_name": "Arcee AI",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.arcee.ai",
   "huggingface": "https://huggingface.co/arcee-ai",
   "github": "https://github.com/arcee-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "San Francisco company training and releasing open models (AFM, Trinity) and tools such as MergeKit and DistillKit.",
   "tagline": "A San Francisco model lab that trains open-weight models and maintains MergeKit, a widely used toolkit for merging open language models.",
   "about": [
    "Arcee AI was founded in 2023 by Mark McQuade, Jacob Solawetz and Brian Benedict to make smaller, specialized language models practical for enterprises. It began by helping teams adapt open models and later moved to training its own models in San Francisco.",
    "Its open tooling includes MergeKit for merging pretrained models, described in a 2024 paper, and DistillKit for distillation. Its current offerings include the Trinity model family, an open source catalog and an API platform.",
    "Investors named on its site include Samsung, M12, Hitachi Ventures and Flybridge. Reporting indicates a 24 million dollar Series A in 2024."
   ],
   "mission": "Arcee states that the future of AI should be something people can build on, adapt and own.",
   "notice": null,
   "known_for": [
    "Trinity open-weight models",
    "MergeKit model merging toolkit",
    "Open source catalog of small models"
   ],
   "focus": [
    "language models",
    "model merging",
    "distillation"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in 2023 to build specialized small language models.",
     "year": "2023"
    },
    {
     "title": "MergeKit paper",
     "note": "The MergeKit toolkit for merging large language models was described in a paper.",
     "year": "2024"
    },
    {
     "title": "Series A funding",
     "note": "Company announced a 24 million dollar Series A.",
     "year": "2024"
    },
    {
     "title": "Trinity manifesto",
     "note": "CTO Lucas Atkins published a case for owning pretraining, ahead of the Trinity model family.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "mergekit",
     "url": "https://github.com/arcee-ai/mergekit",
     "note": "Tools for merging pretrained LLMs; most-starred repo."
    },
    {
     "name": "DistillKit",
     "url": "https://github.com/arcee-ai/DistillKit",
     "note": "LLM distillation toolkit."
    },
    {
     "name": "fastmlx",
     "url": "https://github.com/arcee-ai/fastmlx",
     "note": "MLX hosting API."
    },
    {
     "name": "DALM",
     "url": "https://github.com/arcee-ai/DALM",
     "note": "Domain-adapted RAG toolkit."
    }
   ],
   "hf_highlights": [
    {
     "name": "Trinity-Mini",
     "url": "https://huggingface.co/arcee-ai/Trinity-Mini",
     "kind": "model",
     "note": "Most downloaded Arcee model."
    },
    {
     "name": "AFM-4.5B",
     "url": "https://huggingface.co/arcee-ai/AFM-4.5B",
     "kind": "model",
     "note": "AFM foundation model."
    },
    {
     "name": "LLama-405B-Logits",
     "url": "https://huggingface.co/datasets/arcee-ai/LLama-405B-Logits",
     "kind": "dataset",
     "note": "Teacher logits for distillation."
    },
    {
     "name": "distilabel-intel-orca-dpo-pairs-binarized",
     "url": "https://huggingface.co/datasets/arcee-ai/distilabel-intel-orca-dpo-pairs-binarized",
     "kind": "dataset",
     "note": "Binarised DPO pairs."
    }
   ],
   "papers": [
    {
     "title": "Arcee's MergeKit: A Toolkit for Merging Large Language Models",
     "url": "https://arxiv.org/abs/2403.13257",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "mergekit",
     "url": "https://github.com/arcee-ai/mergekit",
     "note": "Model merging"
    },
    {
     "name": "DistillKit",
     "url": "https://github.com/arcee-ai/DistillKit",
     "note": "Distillation"
    },
    {
     "name": "Trinity-Mini",
     "url": "https://huggingface.co/arcee-ai/Trinity-Mini",
     "note": "Open model"
    }
   ],
   "openness": "Open weights and permissive code.",
   "key_people": [
    {
     "name": "Mark McQuade",
     "role": "Co-founder and CEO",
     "url": "https://yespress.io/mark-mcquade",
     "note": "Previously an early commercial hire at Hugging Face."
    },
    {
     "name": "Lucas Atkins",
     "role": "CTO",
     "url": "https://www.arcee.ai/about/about-us",
     "note": "Company About page names him as CTO."
    },
    {
     "name": "Jacob Solawetz",
     "role": "Co-founder",
     "url": "https://www.arcee.ai/blog/arcee-ai-secures-24m-series-a-to-transform-the-landscape-of-small-language-models",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.arcee.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.arcee.ai",
    "https://huggingface.co/arcee-ai",
    "https://github.com/arcee-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "arcee-ai",
   "mlat": 37.66599,
   "mlon": -122.35565,
   "logo": true
  },
  {
   "orig_id": "audeering",
   "name": "audEERING",
   "alias": null,
   "full_name": "audEERING",
   "type": "startup",
   "founded": 2012,
   "country": "Germany",
   "city": "Gilching",
   "lat": 48.1086,
   "lon": 11.3706,
   "website": "https://www.audeering.com",
   "huggingface": "https://huggingface.co/audeering",
   "github": "https://github.com/audeering",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Audio AI company near Munich, maker of the open-source openSMILE toolkit and widely used speech emotion and age-gender models.",
   "tagline": "Munich-area audio AI company and maker of the openSMILE toolkit and open speech emotion and age-gender models.",
   "about": [
    "audEERING was founded in 2012 as a spin-off of Technical University of Munich research, with Dagmar Schuller, Florian Eyben, Björn Schuller and Martin Wöllmer named as founders in a 2012 report. It is based in Gilching near Munich.",
    "openSMILE began in 2008 at TUM within the EU SEMAINE project. In 2013 audEERING obtained the rights and published version 2.0 under a research licence. The toolkit is free for research and personal use, while commercial users need a licence, so it is source-available rather than fully open.",
    "The company also maintains open repositories on audio machine learning, including the w2v2-how-to guide and wav2vec2-based emotion and age-gender models on Hugging Face."
   ],
   "mission": "audEERING develops audio analysis and speech emotion technology and shares research tooling with the scientific community while selling commercial products.",
   "notice": null,
   "known_for": [
    "openSMILE feature extraction",
    "Speech emotion recognition models",
    "Age and gender audio models",
    "TUM spin-off audio AI"
   ],
   "focus": [
    "speech emotion recognition",
    "audio analysis"
   ],
   "contributions": [
    {
     "title": "openSMILE begins",
     "note": "Eyben, Wöllmer and Schuller started openSMILE at TUM within the SEMAINE project.",
     "year": "2008"
    },
    {
     "title": "Company founded",
     "note": "audEERING was founded in Munich as a TUM spin-off.",
     "year": "2012"
    },
    {
     "title": "openSMILE 2.0",
     "note": "The company acquired the code rights and released version 2.0 under a research licence.",
     "year": "2013"
    },
    {
     "title": "opensmile Python repository",
     "note": "A new openSMILE repository was created on the audeering GitHub organization.",
     "year": "2020"
    },
    {
     "title": "w2v2-how-to",
     "note": "Published a guide and code for wav2vec2 emotion work.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "opensmile",
     "url": "https://github.com/audeering/opensmile",
     "note": "The Munich Open-Source Large-Scale Multimedia Feature Extractor (859 stars)"
    },
    {
     "name": "w2v2-how-to",
     "url": "https://github.com/audeering/w2v2-how-to",
     "note": "How to use our public wav2vec2 dimensional emotion model (564 stars)"
    },
    {
     "name": "opensmile-python",
     "url": "https://github.com/audeering/opensmile-python",
     "note": "Python package for openSMILE (337 stars)"
    },
    {
     "name": "audtorch",
     "url": "https://github.com/audeering/audtorch",
     "note": "Utils and data sets for audio and PyTorch (87 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim",
     "url": "https://huggingface.co/audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim",
     "kind": "model",
     "note": "Model with 177 likes on the Hub."
    },
    {
     "name": "audeering/wav2vec2-large-robust-24-ft-age-gender",
     "url": "https://huggingface.co/audeering/wav2vec2-large-robust-24-ft-age-gender",
     "kind": "model",
     "note": "Model with 88 likes on the Hub."
    },
    {
     "name": "audeering/speech-analysis",
     "url": "https://huggingface.co/spaces/audeering/speech-analysis",
     "kind": "space",
     "note": "Space with 4 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Dawn of the transformer era in speech emotion recognition: closing the valence gap",
     "url": "https://arxiv.org/abs/2203.07378",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "openSMILE",
     "url": "https://github.com/audeering/opensmile",
     "note": "Open-source multimedia feature extractor"
    },
    {
     "name": "wav2vec2 emotion model",
     "url": "https://huggingface.co/audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim",
     "note": "Dimensional speech emotion model"
    },
    {
     "name": "wav2vec2 age-gender model",
     "url": "https://huggingface.co/audeering/wav2vec2-large-robust-24-ft-age-gender",
     "note": "Age and gender estimation model"
    },
    {
     "name": "audtorch",
     "url": "https://github.com/audeering/audtorch",
     "note": "Audio utilities and datasets for PyTorch"
    }
   ],
   "openness": "Commercial company with a large public footprint: tools and several models are open on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Dagmar Schuller",
     "role": "Co-founder and CEO",
     "url": "https://innovationorigins.com/speech-recognition-startup-follows-the-legacy-of-star-trek/",
     "note": null
    },
    {
     "name": "Florian Eyben",
     "role": "Co-founder, openSMILE author",
     "url": "https://en.wikipedia.org/wiki/OpenSMILE",
     "note": null
    },
    {
     "name": "Björn Schuller",
     "role": "Co-founder, openSMILE author",
     "url": "https://en.wikipedia.org/wiki/OpenSMILE",
     "note": null
    },
    {
     "name": "Felix Burkhardt",
     "role": "Researcher",
     "url": "https://github.com/felixbur",
     "note": null
    },
    {
     "name": "Hagen Wierstorf",
     "role": "Researcher",
     "url": "https://hagenw.github.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://www.audeering.com",
    "https://huggingface.co/audeering",
    "https://github.com/audeering",
    "https://arxiv.org/abs/2203.07378"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "audeering",
   "logo": true
  },
  {
   "orig_id": "automl-freiburg",
   "name": "AutoML Groups",
   "alias": "Freiburg and Hannover",
   "full_name": "AutoML Groups (Freiburg and Hannover)",
   "type": "academic",
   "founded": null,
   "country": "Germany",
   "city": "Freiburg",
   "lat": 47.999,
   "lon": 7.8421,
   "website": "https://www.automl.org",
   "huggingface": null,
   "github": "https://github.com/automl",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "AutoML research groups at the Universities of Freiburg and Hannover, creators of auto-sklearn, Auto-PyTorch and SMAC3. The shared automl.org site was closed in August 2026.",
   "tagline": "The AutoML groups of Freiburg and Hannover, creators of auto-sklearn, SMAC3 and Auto-PyTorch.",
   "about": [
    "The AutoML groups at the University of Freiburg and Leibniz University Hannover were led by Frank Hutter and Marius Lindauer, with Matthias Feurer and Katharina Eggensperger among close collaborators. The GitHub organization dates from 2014.",
    "Their system auto-sklearn began as a competition entry and won the first international AutoML challenge in 2015 to 2016 and the second in 2017 to 2018. The 2015 NeurIPS paper introduced it, and auto-sklearn 2.0 appeared in JMLR in 2022. SMAC3 and Auto-PyTorch extend the work to Bayesian optimisation and deep learning.",
    "In an August 2026 post the group announced the closure of automl.org after a decade. Hutter now leads Prior Labs and Lindauer's Hannover group has broadened beyond AutoML."
   ],
   "mission": "The groups aimed to make machine learning accessible by automating model and hyperparameter selection, and released their tools as open source for the community.",
   "notice": null,
   "known_for": [
    "auto-sklearn",
    "SMAC3 Bayesian optimisation",
    "Auto-PyTorch",
    "AutoML challenge wins"
   ],
   "focus": [
    "AutoML"
   ],
   "contributions": [
    {
     "title": "auto-sklearn",
     "note": "The NeurIPS paper Efficient and Robust Automated Machine Learning introduced the system.",
     "year": "2015"
    },
    {
     "title": "First AutoML challenge win",
     "note": "The team won the first international AutoML challenge.",
     "year": "2016"
    },
    {
     "title": "Second AutoML challenge win",
     "note": "PoSH-Auto-sklearn won the second challenge against 41 other systems.",
     "year": "2018"
    },
    {
     "title": "Auto-PyTorch",
     "note": "Released automated deep learning tooling (arXiv 2006.13799).",
     "year": "2020"
    },
    {
     "title": "SMAC3",
     "note": "Published a versatile Bayesian optimisation package (arXiv 2109.09831).",
     "year": "2021"
    },
    {
     "title": "auto-sklearn 2.0",
     "note": "The JMLR paper reported stronger results in 10 minutes than version 1.0 in 60.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "auto-sklearn",
     "url": "https://github.com/automl/auto-sklearn",
     "note": "Automated Machine Learning with scikit-learn (8,148 stars)"
    },
    {
     "name": "Auto-PyTorch",
     "url": "https://github.com/automl/Auto-PyTorch",
     "note": "Automatic architecture search and hyperparameter optimization for PyTorch (2,541 stars)"
    },
    {
     "name": "SMAC3",
     "url": "https://github.com/automl/SMAC3",
     "note": "SMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization (1,273 stars)"
    },
    {
     "name": "HpBandSter",
     "url": "https://github.com/automl/HpBandSter",
     "note": "a distributed Hyperband implementation on Steroids (631 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "SMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization",
     "url": "https://arxiv.org/abs/2109.09831",
     "year": "2021"
    },
    {
     "title": "Auto-PyTorch Tabular: Multi-Fidelity MetaLearning for Efficient and Robust AutoDL",
     "url": "https://arxiv.org/abs/2006.13799",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "auto-sklearn",
     "url": "https://github.com/automl/auto-sklearn",
     "note": "Automated machine learning with scikit-learn (8,000+ stars)"
    },
    {
     "name": "Auto-PyTorch",
     "url": "https://github.com/automl/Auto-PyTorch",
     "note": "Architecture search and hyperparameter optimisation for PyTorch"
    },
    {
     "name": "SMAC3",
     "url": "https://github.com/automl/SMAC3",
     "note": "Bayesian optimisation package for hyperparameters"
    },
    {
     "name": "HpBandSter",
     "url": "https://github.com/automl/HpBandSter",
     "note": "Distributed Hyperband implementation"
    }
   ],
   "openness": "Libraries are open source on GitHub under the automl organisation.",
   "key_people": [
    {
     "name": "Frank Hutter",
     "role": "Co-lead, now co-founder and CEO of Prior Labs",
     "url": "https://automl.org/",
     "note": null
    },
    {
     "name": "Marius Lindauer",
     "role": "Professor at Leibniz University Hannover",
     "url": "https://www.alphaxiv.org/researchers/marius-lindauer",
     "note": null
    },
    {
     "name": "Matthias Feurer",
     "role": "Lead author of the auto-sklearn papers",
     "url": "https://automl.org/",
     "note": null
    },
    {
     "name": "Katharina Eggensperger",
     "role": "Co-author of auto-sklearn and SMAC3 work",
     "url": "https://automl.org/",
     "note": null
    },
    {
     "name": "Timur Carstensen",
     "role": "Member of the automl GitHub organization",
     "url": "https://github.com/timurcarstensen",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.automl.org",
    "https://github.com/automl",
    "https://arxiv.org/abs/2109.09831"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "automl-groups",
   "mlat": 47.9557,
   "mlon": 7.8421,
   "logo": true
  },
  {
   "orig_id": "baai",
   "name": "BAAI",
   "alias": "Beijing Academy of Artificial Intelligence",
   "full_name": "BAAI (Beijing Academy of Artificial Intelligence)",
   "type": "nonprofit",
   "founded": 2018,
   "country": "China",
   "city": "Beijing",
   "lat": 39.9,
   "lon": 116.41,
   "website": "https://www.baai.ac.cn/english.html",
   "huggingface": "https://huggingface.co/BAAI",
   "github": "https://github.com/FlagOpen",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Describes itself as a private non-profit AI research and development organisation in Beijing. Publishes models and datasets on Hugging Face.",
   "tagline": "A Beijing nonprofit AI research academy behind the Wudao models, the FlagOpen ecosystem and the BGE embedding models.",
   "about": [
    "The Beijing Academy of Artificial Intelligence was founded in November 2018, with Zhang Hongjiang listed as its founder.",
    "It announced the Wudao 2.0 pretrained model system, and it maintains the FlagAI toolkit and the FlagOpen open-source ecosystem. It also publishes the BGE embedding models and the Emu3 multimodal models on Hugging Face."
   ],
   "mission": "BAAI aims to advance foundational AI research in China and share models, data and tools with the wider community.",
   "notice": null,
   "known_for": [
    "Wudao large-scale pretrained models",
    "BGE embedding models",
    "FlagOpen ecosystem",
    "Emu3 multimodal models"
   ],
   "focus": [
    "datasets",
    "language models",
    "embodied AI",
    "Chinese"
   ],
   "contributions": [
    {
     "title": "Academy founded",
     "note": "BAAI was established in Beijing in November 2018.",
     "year": "2018"
    },
    {
     "title": "Wudao 2.0 announced",
     "note": "A large pretrained model system announced in 2022.",
     "year": "2022"
    },
    {
     "title": "FlagAI joins LF AI & Data",
     "note": "The FlagAI toolkit was approved as a sandbox project of the Linux Foundation's AI foundation.",
     "year": "2022"
    },
    {
     "title": "Emu3 and BGE papers",
     "note": "Published the Emu3 multimodal paper (2024) and the C-Pack paper behind BGE embeddings (2023).",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "FlagEmbedding",
     "url": "https://github.com/FlagOpen/FlagEmbedding",
     "note": "Retrieval and retrieval-augmented LLMs."
    },
    {
     "name": "RoboBrain2.5",
     "url": "https://github.com/FlagOpen/RoboBrain2.5",
     "note": "Embodied brain model."
    },
    {
     "name": "Robo-Dopamine",
     "url": "https://github.com/FlagOpen/Robo-Dopamine",
     "note": "Process reward modeling for robotic manipulation."
    },
    {
     "name": "RoboOS",
     "url": "https://github.com/FlagOpen/RoboOS",
     "note": "Embodied operating system for multi-robot collaboration."
    }
   ],
   "hf_highlights": [
    {
     "name": "bge-small-en-v1.5",
     "url": "https://huggingface.co/BAAI/bge-small-en-v1.5",
     "kind": "model",
     "note": "Very widely used English embedding model."
    },
    {
     "name": "bge-m3",
     "url": "https://huggingface.co/BAAI/bge-m3",
     "kind": "model",
     "note": "Multilingual, multi-granularity embedding model."
    },
    {
     "name": "bge-reranker-v2-m3",
     "url": "https://huggingface.co/BAAI/bge-reranker-v2-m3",
     "kind": "model",
     "note": "Multilingual reranker."
    },
    {
     "name": "Infinity-MM",
     "url": "https://huggingface.co/datasets/BAAI/Infinity-MM",
     "kind": "dataset",
     "note": "Large multimodal instruction dataset."
    }
   ],
   "papers": [
    {
     "title": "Emu3: Next-Token Prediction is All You Need",
     "url": "https://arxiv.org/abs/2409.18869",
     "year": "2024"
    },
    {
     "title": "M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation",
     "url": "https://arxiv.org/abs/2402.03216",
     "year": "2024"
    },
    {
     "title": "C-Pack: Packed Resources For General Chinese Embeddings",
     "url": "https://arxiv.org/abs/2309.07597",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "CS-Dialogue",
     "url": "https://huggingface.co/datasets/BAAI/CS-Dialogue",
     "note": "Dataset"
    },
    {
     "name": "RealTalk-CN",
     "url": "https://huggingface.co/datasets/BAAI/RealTalk-CN",
     "note": "Dataset"
    }
   ],
   "openness": "Models and datasets on Hugging Face; licences vary per release.",
   "key_people": [
    {
     "name": "Zhang Hongjiang",
     "role": "Founder",
     "url": "https://en.wikipedia.org/wiki/Zhang_Hongjiang",
     "note": null
    },
    {
     "name": "Huang Tiejun",
     "role": "Chairman of the Board; founding president",
     "url": "https://www.geopolitechs.org/p/baai-chairman-huang-tiejun-the-path",
     "note": "Led the academy as its first dean and president."
    },
    {
     "name": "Wang Zhongyuan",
     "role": "President",
     "url": "https://syncedreview.com/2024/06/17/ai-pioneers-gather-at-baai-2024-unveiling-innovations-in-large-scaled-ai-models-for-language-multimodal-embodied-bio-computing-and-flagopen-2-0/",
     "note": "Reported to have succeeded Huang as president in 2024."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the institute's website or the FlagOpen GitHub organisation.",
   "sources": [
    "https://huggingface.co/BAAI",
    "https://www.baai.ac.cn/english.html"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "baai",
   "mlat": 39.8441,
   "mlon": 116.41,
   "logo": true
  },
  {
   "orig_id": "bespoke-labs",
   "name": "Bespoke Labs",
   "alias": null,
   "full_name": "Bespoke Labs",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "Mountain View",
   "lat": 37.3861,
   "lon": -122.0839,
   "website": "https://bespokelabs.ai",
   "huggingface": "https://huggingface.co/bespokelabs",
   "github": "https://github.com/bespokelabsai",
   "x": "https://x.com/bespokelabsai",
   "discord": "https://discord.com/invite/KqpXvpzVBS",
   "linkedin": null,
   "contact_email": null,
   "summary": "Startup building tools and open models for data curation and verifiable fine-tuning, and a lead team behind the OpenThoughts project.",
   "tagline": "A data research lab that builds the Curator data tooling, RL environments and benchmarks, and co-leads the OpenThoughts open reasoning data project.",
   "about": [
    "Bespoke Labs is a data research lab in the SF Bay Area. Its focus is RL environments, agent evaluation and optimisation, and benchmarking for production agents.",
    "It maintains Curator, an open library for data curation, and contributes to Terminal-Bench and GEPA. Its site lists OpenThoughts, an open reasoning dataset presented at ICLR 2026.",
    "The OpenThoughts collaboration, led by researchers across Stanford, UC Berkeley, UT Austin, LAION and other institutions, launched in January 2025 and released OpenThoughts3 in June 2025."
   ],
   "mission": "Bespoke aims to give frontier labs and enterprises the data and environments needed to train long-horizon agents.",
   "notice": null,
   "known_for": [
    "OpenThoughts open reasoning datasets",
    "Curator data curation library",
    "Terminal-Bench agent benchmark"
   ],
   "focus": [
    "synthetic data",
    "post-training",
    "fact-checking",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in 2024 by Mahesh Sathiamoorthy and Alex Dimakis.",
     "year": "2024"
    },
    {
     "title": "OpenThoughts launched",
     "note": "The collaboration started publishing open reasoning data.",
     "year": "2025"
    },
    {
     "title": "OpenThinker-32B released",
     "note": "An open reasoning model from the project.",
     "year": "2025"
    },
    {
     "title": "OpenThoughts3 released",
     "note": "Described as the project's best dataset recipe; the OpenThoughts paper appeared in June 2025.",
     "year": "2025"
    },
    {
     "title": "Terminal-Bench and GEPA",
     "note": "Contributions to the Terminal-Bench agent benchmark and the GEPA prompt optimizer, both accepted to ICLR 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "curator",
     "url": "https://github.com/bespokelabsai/curator",
     "note": "Synthetic data curation library"
    },
    {
     "name": "verifiers",
     "url": "https://github.com/bespokelabsai/verifiers",
     "note": "Verifiers for LLM reinforcement learning"
    },
    {
     "name": "nimble",
     "url": "https://github.com/bespokelabsai/nimble",
     "note": "Typed local decisions and model evaluation"
    }
   ],
   "hf_highlights": [
    {
     "name": "Bespoke-MiniCheck-7B",
     "url": "https://huggingface.co/bespokelabs/Bespoke-MiniCheck-7B",
     "kind": "model",
     "note": "Fact-checking model"
    },
    {
     "name": "Bespoke-Stratos-17k",
     "url": "https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-17k",
     "kind": "dataset",
     "note": "Reasoning dataset"
    },
    {
     "name": "Bespoke-Stratos-7B",
     "url": "https://huggingface.co/bespokelabs/Bespoke-Stratos-7B",
     "kind": "model",
     "note": "Distilled reasoning model"
    }
   ],
   "papers": [
    {
     "title": "OpenThoughts: Data Recipes for Reasoning Models",
     "url": "https://arxiv.org/abs/2506.04178",
     "year": "2025"
    },
    {
     "title": "GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning",
     "url": "https://arxiv.org/abs/2507.19457",
     "year": "2025"
    },
    {
     "title": "MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents",
     "url": "https://arxiv.org/abs/2404.10774",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "curator",
     "url": "https://github.com/bespokelabsai/curator",
     "note": "Synthetic data curation library"
    },
    {
     "name": "verifiers",
     "url": "https://github.com/bespokelabsai/verifiers",
     "note": "Verifiers for LLM reinforcement learning"
    },
    {
     "name": "nimble",
     "url": "https://github.com/bespokelabsai/nimble",
     "note": "Typed local decisions and model evaluation"
    }
   ],
   "openness": "Apache-2.0 Curator library; open weights and datasets such as Bespoke-Stratos and MiniCheck on Hugging Face.",
   "key_people": [
    {
     "name": "Mahesh Sathiamoorthy",
     "role": "Co-founder and CEO",
     "url": "https://www.ai.engineer/speakers/mahesh-sathiamoorthy",
     "note": "Previously at Google DeepMind."
    },
    {
     "name": "Alex Dimakis",
     "role": "Co-founder and Chief Scientist; professor at UC Berkeley",
     "url": "https://people.eecs.berkeley.edu/~alexdimakis/",
     "note": null
    },
    {
     "name": "Negin Raoof",
     "role": "Researcher, OpenThoughts co-author (UC Berkeley)",
     "url": "https://arxiv.org/abs/2506.04178",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the website.",
   "contact_note": "Via website.",
   "sources": [
    "https://bespokelabs.ai/",
    "https://github.com/bespokelabsai",
    "https://huggingface.co/bespokelabs"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "bespoke-labs",
   "mlat": 37.3428,
   "mlon": -122.0839,
   "logo": true
  },
  {
   "orig_id": "bethgelab",
   "name": "Bethge Lab",
   "alias": "Tuebingen",
   "full_name": "Bethge Lab (Tuebingen)",
   "type": "academic",
   "founded": 2014,
   "country": "Germany",
   "city": "Tubingen",
   "lat": 48.5216,
   "lon": 9.0576,
   "website": "https://bethgelab.org",
   "huggingface": "https://huggingface.co/bethgelab",
   "github": "https://github.com/bethgelab",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Bethge Lab at the University of Tuebingen studies robustness and human-like vision, releasing Foolbox and several robustness benchmarks.",
   "tagline": "Tübingen lab studying robustness and human-like vision, known for the Foolbox toolbox and open robustness benchmarks.",
   "about": [
    "The Bethge Lab is led by Matthias Bethge, Professor of Computational Neuroscience and Machine Learning at the University of Tübingen and director of the Tübingen AI Center. Its GitHub organization, created in 2014, carries the tagline Perceiving Neural Networks.",
    "The lab released Foolbox in 2017, a Python toolbox for benchmarking adversarial robustness under an MIT licence, and imagecorruptions and a detection robustness benchmark in 2019. It also published work on ImageNet-C robustness and on comparing machine and human vision.",
    "Wieland Brendel, Jonas Rauber and Robert Geirhos appear among its public GitHub members and papers, and its code serves vision and robustness researchers."
   ],
   "mission": "The lab seeks to understand how neural networks perceive images and to make them robust and more human-like, releasing benchmarks and tools openly.",
   "notice": null,
   "known_for": [
    "Foolbox adversarial toolbox",
    "Corruption robustness benchmarks",
    "Human versus machine vision",
    "ImageNet-C robustness training"
   ],
   "focus": [
    "robustness",
    "vision"
   ],
   "contributions": [
    {
     "title": "Foolbox",
     "note": "Released an adversarial robustness toolbox (arXiv 1707.04131).",
     "year": "2017"
    },
    {
     "title": "Robustness benchmarks",
     "note": "Released the imagecorruptions package and a detection robustness benchmark (arXiv 1907.07484).",
     "year": "2019"
    },
    {
     "title": "Noise training baseline",
     "note": "Rusak et al. showed Gaussian and Speckle noise training is a strong ImageNet-C baseline (arXiv 2001.06057).",
     "year": "2020"
    },
    {
     "title": "Human versus machine vision",
     "note": "Released an open benchmark comparing models to humans (arXiv 2106.07411).",
     "year": "2021"
    }
   ],
   "github_highlights": [
    {
     "name": "foolbox",
     "url": "https://github.com/bethgelab/foolbox",
     "note": "A Python toolbox to create adversarial examples that fool neural networks in PyTorch, TensorFlow, and JAX (2,984 stars)"
    },
    {
     "name": "imagecorruptions",
     "url": "https://github.com/bethgelab/imagecorruptions",
     "note": "Python package to corrupt arbitrary images. (486 stars)"
    },
    {
     "name": "model-vs-human",
     "url": "https://github.com/bethgelab/model-vs-human",
     "note": "Benchmark your model on out-of-distribution datasets with carefully collected human comparison data (NeurIPS 2 (365 stars)"
    },
    {
     "name": "siamese-mask-rcnn",
     "url": "https://github.com/bethgelab/siamese-mask-rcnn",
     "note": "Siamese Mask R-CNN model for one-shot instance segmentation (354 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "bethgelab/ViT-B-16-SigLIP-256_cabs-fm_recap_0.8",
     "url": "https://huggingface.co/bethgelab/ViT-B-16-SigLIP-256_cabs-fm_recap_0.8",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "bethgelab/motion_energy_segmentation_model",
     "url": "https://huggingface.co/bethgelab/motion_energy_segmentation_model",
     "kind": "model",
     "note": "Model with 0 likes on the Hub."
    },
    {
     "name": "bethgelab/CuratedThoughts",
     "url": "https://huggingface.co/datasets/bethgelab/CuratedThoughts",
     "kind": "dataset",
     "note": "Dataset with 45 likes on the Hub."
    },
    {
     "name": "bethgelab/lm-similarity",
     "url": "https://huggingface.co/spaces/bethgelab/lm-similarity",
     "kind": "space",
     "note": "Space with 17 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Foolbox: A Python toolbox to benchmark the robustness of machine learning models",
     "url": "https://arxiv.org/abs/1707.04131",
     "year": "2017"
    },
    {
     "title": "Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming",
     "url": "https://arxiv.org/abs/1907.07484",
     "year": "2019"
    },
    {
     "title": "Partial success in closing the gap between human and machine vision",
     "url": "https://arxiv.org/abs/2106.07411",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Foolbox",
     "url": "https://github.com/bethgelab/foolbox",
     "note": "Toolbox for adversarial examples (about 3,000 stars)"
    },
    {
     "name": "imagecorruptions",
     "url": "https://github.com/bethgelab/imagecorruptions",
     "note": "Image corruption package for robustness tests"
    },
    {
     "name": "model-vs-human",
     "url": "https://github.com/bethgelab/model-vs-human",
     "note": "Benchmark of models against human behaviour"
    },
    {
     "name": "CuratedThoughts",
     "url": "https://huggingface.co/datasets/bethgelab/CuratedThoughts",
     "note": "Dataset on the Hub"
    }
   ],
   "openness": "Code and datasets are open on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Matthias Bethge",
     "role": "Professor and lab head, director of the Tübingen AI Center",
     "url": "https://bethgelab.org",
     "note": null
    },
    {
     "name": "Wieland Brendel",
     "role": "Researcher, Foolbox and robustness papers",
     "url": "http://www.wielandbrendel.de",
     "note": null
    },
    {
     "name": "Jonas Rauber",
     "role": "Foolbox lead author",
     "url": "https://jonasrauber.de/",
     "note": null
    },
    {
     "name": "Robert Geirhos",
     "role": "Researcher on shape and texture bias",
     "url": "https://robertgeirhos.com",
     "note": null
    },
    {
     "name": "Claudio Michaelis",
     "role": "Detection robustness benchmark author",
     "url": "https://github.com/michaelisc",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://bethgelab.org",
    "https://huggingface.co/bethgelab",
    "https://github.com/bethgelab",
    "https://arxiv.org/abs/1707.04131"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "bethge-lab",
   "logo": true
  },
  {
   "orig_id": "bharatgen",
   "name": "BharatGen",
   "alias": null,
   "full_name": "BharatGen",
   "type": "institute",
   "founded": null,
   "country": "India",
   "city": "Mumbai",
   "lat": 19.08,
   "lon": 72.88,
   "website": "https://bharatgen.com/",
   "huggingface": "https://huggingface.co/bharatgenai",
   "github": "https://github.com/Bharatgen-Tech",
   "x": "https://twitter.com/BharatGen_Com",
   "discord": null,
   "linkedin": "https://linkedin.com/company/bharatgen",
   "contact_email": null,
   "summary": "India-focused consortium releasing language, speech and vision models plus the BhashaBench evaluation suite. Government-funded.",
   "tagline": "An India-wide consortium anchored at IIT Bombay that releases open-weight text, speech and vision models for Indian languages.",
   "about": [
    "BharatGen is a Section 8 not-for-profit company, the BharatGen Technology Foundation, anchored at IIT Bombay and drawing on nine consortium institutions. It is backed by India's Department of Science and Technology and the IndiaAI Mission.",
    "A blog post dates its launch to October 2024. Its models include Param2, a 17B text model, the Sooktam text-to-speech and Shrutam speech-to-text families, and the Patram vision-language models."
   ],
   "mission": "Its stated slogan is built from scratch, open weights, made for India.",
   "notice": null,
   "known_for": [
    "Param open-weight Indian language models",
    "Sooktam speech synthesis and Shrutam recognition",
    "Patram vision-language models"
   ],
   "focus": [
    "language models",
    "evaluation",
    "Indic languages"
   ],
   "contributions": [
    {
     "title": "First models demonstrated",
     "note": "Foundational models were announced at the India Mobile Congress on October 17, 2024.",
     "year": "2024"
    },
    {
     "title": "BharatGen launched",
     "note": "The programme was launched in New Delhi.",
     "year": "2024"
    },
    {
     "title": "Param-1 released",
     "note": "A 2.9 billion parameter bilingual base model described as trained from scratch.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "Tables-Decoded",
     "url": "https://github.com/Bharatgen-Tech/Tables-Decoded",
     "note": "Table understanding repo."
    },
    {
     "name": "syntheticgen",
     "url": "https://github.com/Bharatgen-Tech/syntheticgen",
     "note": "Synthetic data generation."
    },
    {
     "name": "BhashaBench",
     "url": "https://github.com/Bharatgen-Tech/BhashaBench",
     "note": "BhashaBench code."
    }
   ],
   "hf_highlights": [
    {
     "name": "bharatgenai/Param2-17B-A2.4B-Thinking",
     "url": "https://huggingface.co/bharatgenai/Param2-17B-A2.4B-Thinking",
     "kind": "model",
     "note": "Most downloaded."
    },
    {
     "name": "bharatgenai/patram-7b-instruct",
     "url": "https://huggingface.co/bharatgenai/patram-7b-instruct",
     "kind": "model",
     "note": "Vision-language model."
    },
    {
     "name": "bharatgenai/Param-1-2.9B-Instruct",
     "url": "https://huggingface.co/bharatgenai/Param-1-2.9B-Instruct",
     "kind": "model",
     "note": "Instruct model."
    },
    {
     "name": "bharatgenai/BhashaBench-Multi",
     "url": "https://huggingface.co/datasets/bharatgenai/BhashaBench-Multi",
     "kind": "dataset",
     "note": "Multi-domain Indic benchmark."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Param2-17B-A2.4B-Thinking",
     "url": "https://huggingface.co/bharatgenai/Param2-17B-A2.4B-Thinking",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Shrutam-2",
     "url": "https://huggingface.co/bharatgenai/Shrutam-2",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "patram-7b-instruct",
     "url": "https://huggingface.co/bharatgenai/patram-7b-instruct",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [
    {
     "name": "Ganesh Ramakrishnan",
     "role": "Professor, IIT Bombay; leads BharatGen",
     "url": "https://www.cse.iitb.ac.in/~ganesh/",
     "note": "Presented the first BharatGen models in October 2024."
    }
   ],
   "discord_note": null,
   "contact_note": "bharatgen.com has a Contact us page.",
   "sources": [
    "https://huggingface.co/bharatgenai",
    "https://bharatgen.com/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "bharatgen",
   "logo": true
  },
  {
   "orig_id": "bigcode",
   "name": "BigCode",
   "alias": null,
   "full_name": "BigCode",
   "type": "collective",
   "founded": 2022,
   "country": "United States",
   "city": "New York",
   "lat": 40.7128,
   "lon": -74.006,
   "website": "https://www.bigcode-project.org",
   "huggingface": "https://huggingface.co/bigcode",
   "github": "https://github.com/bigcode-project",
   "x": "https://twitter.com/BigCodeProject",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "BigCode is an open scientific collaboration, run by Hugging Face and ServiceNow Research, on code language models. It released the StarCoder models, The Stack datasets and evaluation tools. Independence is partial because two companies host it.",
   "tagline": "Open scientific collaboration run by Hugging Face and ServiceNow Research that released The Stack and the StarCoder family of code models.",
   "about": [
    "BigCode was established in September 2022 as an open scientific collaboration on the responsible development of code language models. It is stewarded by Hugging Face and ServiceNow and has drawn more than 1,100 members.",
    "The Stack is a permissively licensed code dataset with an opt-out process for developers. StarCoder, trained on 1 trillion tokens, followed in May 2023 and StarCoder2 in early 2024. The project also maintains an evaluation harness.",
    "Leandro von Werra of Hugging Face and Harm de Vries of ServiceNow Research co-lead the project. Because two companies host it, its independence is partial."
   ],
   "mission": "BigCode aims to develop code language models openly and responsibly, with transparent data governance, community participation and reproducible evaluation.",
   "notice": null,
   "known_for": [
    "StarCoder code models",
    "The Stack dataset with opt-out",
    "Open data governance",
    "Code evaluation harness"
   ],
   "focus": [
    "language models",
    "code",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Project established",
     "note": "BigCode began as an open scientific collaboration.",
     "year": "2022"
    },
    {
     "title": "The Stack",
     "note": "Released a large permissively licensed code dataset with opt-out governance.",
     "year": "2022"
    },
    {
     "title": "StarCoder",
     "note": "Released a model trained on 1 trillion tokens from The Stack.",
     "year": "2023"
    },
    {
     "title": "StarCoder2",
     "note": "Released the second generation of models (arXiv 2402.19173).",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "starcoder",
     "url": "https://github.com/bigcode-project/starcoder",
     "note": "Home of StarCoder fine-tuning and inference code."
    },
    {
     "name": "starcoder2",
     "url": "https://github.com/bigcode-project/starcoder2",
     "note": "Home of StarCoder2."
    },
    {
     "name": "bigcode-evaluation-harness",
     "url": "https://github.com/bigcode-project/bigcode-evaluation-harness",
     "note": "Evaluation framework for code generation models."
    },
    {
     "name": "bigcodebench",
     "url": "https://github.com/bigcode-project/bigcodebench",
     "note": "BigCodeBench benchmark (ICLR 2025)."
    }
   ],
   "hf_highlights": [
    {
     "name": "starcoder2-3b",
     "url": "https://huggingface.co/bigcode/starcoder2-3b",
     "kind": "model",
     "note": "StarCoder2 at 3B."
    },
    {
     "name": "starcoder2-15b",
     "url": "https://huggingface.co/bigcode/starcoder2-15b",
     "kind": "model",
     "note": "StarCoder2 flagship 15B."
    },
    {
     "name": "bigcodebench",
     "url": "https://huggingface.co/datasets/bigcode/bigcodebench",
     "kind": "dataset",
     "note": "BigCodeBench evaluation data."
    },
    {
     "name": "the-stack",
     "url": "https://huggingface.co/datasets/bigcode/the-stack",
     "kind": "dataset",
     "note": "The Stack source-code dataset."
    }
   ],
   "papers": [
    {
     "title": "StarCoder: may the source be with you!",
     "url": "https://arxiv.org/abs/2305.06161",
     "year": "2023"
    },
    {
     "title": "StarCoder 2 and The Stack v2: The Next Generation",
     "url": "https://arxiv.org/abs/2402.19173",
     "year": "2024"
    },
    {
     "title": "SantaCoder: don't reach for the stars!",
     "url": "https://arxiv.org/abs/2301.03988",
     "year": "2023"
    },
    {
     "title": "OctoPack: Instruction Tuning Code Large Language Models",
     "url": "https://arxiv.org/abs/2308.07124",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "StarCoder2",
     "url": "https://huggingface.co/bigcode/starcoder2-3b",
     "note": "Open code model"
    },
    {
     "name": "The Stack",
     "url": "https://huggingface.co/datasets/bigcode/the-stack",
     "note": "Code dataset"
    },
    {
     "name": "BigCodeBench",
     "url": "https://huggingface.co/datasets/bigcode/bigcodebench",
     "note": "Benchmark"
    },
    {
     "name": "evaluation harness",
     "url": "https://github.com/bigcode-project/bigcode-evaluation-harness",
     "note": "Evaluation framework"
    }
   ],
   "openness": "Open weights under the BigCode OpenRAIL-M family, with datasets and code public.",
   "key_people": [
    {
     "name": "Leandro von Werra",
     "role": "Co-lead, Hugging Face",
     "url": "https://github.com/lvwerra",
     "note": null
    },
    {
     "name": "Harm de Vries",
     "role": "Co-lead, ServiceNow Research",
     "url": "https://arxiv.org/abs/2305.06161",
     "note": null
    },
    {
     "name": "Loubna Ben Allal",
     "role": "Contributor, StarCoder",
     "url": "https://github.com/loubnabnl",
     "note": null
    },
    {
     "name": "Arjun Guha",
     "role": "Contributor to BigCode repositories",
     "url": "https://github.com/arjunguha",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.bigcode-project.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.bigcode-project.org",
    "https://huggingface.co/bigcode",
    "https://github.com/bigcode-project"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "bigcode",
   "mlat": 40.6569,
   "mlon": -74.006,
   "logo": true
  },
  {
   "orig_id": "bigscience",
   "name": "BigScience",
   "alias": null,
   "full_name": "BigScience",
   "type": "collective",
   "founded": 2021,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://bigscience.huggingface.co",
   "huggingface": "https://huggingface.co/bigscience",
   "github": "https://github.com/bigscience-workshop",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "BigScience was a year-long open research workshop that produced the BLOOM multilingual language model and many related datasets and tools. Hosting is on Hugging Face, and the project is now mainly an archive. City is a placeholder for a distributed collaboration.",
   "tagline": "A year-long open research workshop coordinated by Hugging Face that produced the 176B-parameter BLOOM multilingual model.",
   "about": [
    "BigScience was a volunteer-driven research workshop in which several hundred researchers worked in the open, with Hugging Face coordinating and French government funding supporting it.",
    "The main output was BLOOM, released on 12 July 2022 with 176 billion parameters, 46 natural languages and 13 programming languages. It was trained between March and July 2022 on the Jean Zay supercomputer on the ROOTS corpus.",
    "The code is Apache 2.0 and the weights use the BigScience Responsible AI License. Related releases include xP3 and BLOOMZ."
   ],
   "mission": "The workshop aimed to study large language models openly, with participants from many countries sharing data, code and evaluation.",
   "notice": null,
   "known_for": [
    "BLOOM 176B multilingual model",
    "ROOTS corpus",
    "xP3 and BLOOMZ instruction tuning"
   ],
   "focus": [
    "language models",
    "datasets",
    "distributed"
   ],
   "contributions": [
    {
     "title": "T0 prompted training",
     "note": "Multitask Prompted Training Enables Zero-Shot Task Generalization was published.",
     "year": "2021"
    },
    {
     "title": "BLOOM training begins",
     "note": "Training ran on the Jean Zay supercomputer from March.",
     "year": "2022"
    },
    {
     "title": "BLOOM released",
     "note": "A 176B open model in 46 languages and 13 programming languages.",
     "year": "2022"
    },
    {
     "title": "xP3 and BLOOMZ released",
     "note": "A multilingual instruction dataset and a model tuned on it.",
     "year": "2022"
    },
    {
     "title": "ROOTS paper",
     "note": "The ROOTS corpus paper was presented at NeurIPS 2022 and posted to arXiv in 2023.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "petals",
     "url": "https://github.com/bigscience-workshop/petals",
     "note": "Run LLMs at home, BitTorrent-style; most-starred repo."
    },
    {
     "name": "promptsource",
     "url": "https://github.com/bigscience-workshop/promptsource",
     "note": "Toolkit for creating and sharing natural-language prompts."
    },
    {
     "name": "Megatron-DeepSpeed",
     "url": "https://github.com/bigscience-workshop/Megatron-DeepSpeed",
     "note": "Training code used for BLOOM."
    },
    {
     "name": "xmtf",
     "url": "https://github.com/bigscience-workshop/xmtf",
     "note": "Code and data for crosslingual multitask finetuning."
    }
   ],
   "hf_highlights": [
    {
     "name": "bloomz-560m",
     "url": "https://huggingface.co/bigscience/bloomz-560m",
     "kind": "model",
     "note": "Most downloaded BLOOMZ checkpoint."
    },
    {
     "name": "bloom-560m",
     "url": "https://huggingface.co/bigscience/bloom-560m",
     "kind": "model",
     "note": "Small BLOOM base model."
    },
    {
     "name": "P3",
     "url": "https://huggingface.co/datasets/bigscience/P3",
     "kind": "dataset",
     "note": "Public Pool of Prompts used to train T0."
    },
    {
     "name": "xP3",
     "url": "https://huggingface.co/datasets/bigscience/xP3",
     "kind": "dataset",
     "note": "Multilingual prompted dataset for BLOOMZ/mT0."
    }
   ],
   "papers": [
    {
     "title": "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model",
     "url": "https://arxiv.org/abs/2211.05100",
     "year": "2022"
    },
    {
     "title": "The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset",
     "url": "https://arxiv.org/abs/2303.03915",
     "year": "2023"
    },
    {
     "title": "Multitask Prompted Training Enables Zero-Shot Task Generalization",
     "url": "https://arxiv.org/abs/2110.08207",
     "year": "2021"
    },
    {
     "title": "Crosslingual Generalization through Multitask Finetuning",
     "url": "https://arxiv.org/abs/2211.01786",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "BLOOM / BLOOMZ",
     "url": "https://huggingface.co/bigscience/bloomz-560m",
     "note": "Multilingual models"
    },
    {
     "name": "P3",
     "url": "https://huggingface.co/datasets/bigscience/P3",
     "note": "Prompt dataset"
    },
    {
     "name": "Petals",
     "url": "https://github.com/bigscience-workshop/petals",
     "note": "Distributed inference"
    },
    {
     "name": "promptsource",
     "url": "https://github.com/bigscience-workshop/promptsource",
     "note": "Prompt toolkit"
    }
   ],
   "openness": "Open weights, datasets and code, with BLOOM under the Responsible AI License.",
   "key_people": [
    {
     "name": "Thomas Wolf",
     "role": "Co-founder, Hugging Face; workshop organizer",
     "url": "https://thomwolf.io/",
     "note": null
    },
    {
     "name": "Teven Le Scao",
     "role": "First named author of the BLOOM paper",
     "url": "https://arxiv.org/abs/2211.05100",
     "note": null
    },
    {
     "name": "Victor Sanh",
     "role": "Working group chair; lead author of T0",
     "url": "https://arxiv.org/abs/2110.08207",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://bigscience.huggingface.co); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co/bigscience",
    "https://github.com/bigscience-workshop",
    "https://arxiv.org/abs/2211.05100"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "bigscience",
   "mlat": 48.76048,
   "mlon": 2.40879,
   "logo": true
  },
  {
   "orig_id": "bioptimus",
   "name": "Bioptimus",
   "alias": null,
   "full_name": "Bioptimus",
   "type": "startup",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.bioptimus.com",
   "huggingface": "https://huggingface.co/bioptimus",
   "github": "https://github.com/bioptimus",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/bioptimus",
   "contact_email": null,
   "summary": "Paris startup building foundation models for biology and pathology. It has released the H-optimus pathology models on Hugging Face.",
   "tagline": "A Paris startup building foundation models for pathology and biology, including the openly released H-Optimus pathology models.",
   "about": [
    "Bioptimus is a French company that describes its goal as a world model of biology. Its platforms are H-Optimus for pathology, M-Optimus for multimodal data and STELA, a multi-institutional data initiative.",
    "It announced H-Optimus-0 in July 2024, a 1.1-billion-parameter pathology model described as open source. It raised a 35 million dollar seed round led by Sofinnova Partners and, in January 2025, 41 million dollars led by Cathay Innovation."
   ],
   "mission": "Bioptimus says it aims to build a world model of biology across modalities and scales.",
   "notice": null,
   "known_for": [
    "H-Optimus pathology foundation models",
    "M-Optimus multimodal model",
    "STELA data initiative"
   ],
   "focus": [
    "science",
    "biology",
    "pathology",
    "vision"
   ],
   "contributions": [
    {
     "title": "Seed round",
     "note": "A 35 million dollar seed round led by Sofinnova Partners.",
     "year": "2024"
    },
    {
     "title": "H-Optimus-0 announced",
     "note": "A 1.1B-parameter pathology foundation model.",
     "year": "2024"
    },
    {
     "title": "Series A funding",
     "note": "A 41M dollar round led by Cathay Innovation.",
     "year": "2025"
    },
    {
     "title": "M-Optimus unveiled",
     "note": "A multimodal model combining histology, RNA sequencing, spatial transcriptomics and clinical data, announced December 17, 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "releases",
     "url": "https://github.com/bioptimus/releases",
     "note": "Open-source repository (111 stars)"
    },
    {
     "name": "h1-jumpstart",
     "url": "https://github.com/bioptimus/h1-jumpstart",
     "note": "Open-source repository (4 stars)"
    },
    {
     "name": "m-jumpstart",
     "url": "https://github.com/bioptimus/m-jumpstart",
     "note": "Getting started with M-Optimus on AWS   SageMaker (2 stars)"
    },
    {
     "name": "dynolog-bioptimus",
     "url": "https://github.com/bioptimus/dynolog-bioptimus",
     "note": "Customized fork of Dynolog. Dynolog is a telemetry daemon for (GPU & CPU) performance monitoring and tracing,  (1 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "bioptimus/H-optimus-0",
     "url": "https://huggingface.co/bioptimus/H-optimus-0",
     "kind": "model",
     "note": "Model: 73,031 downloads, 121 likes."
    },
    {
     "name": "bioptimus/H-optimus-1",
     "url": "https://huggingface.co/bioptimus/H-optimus-1",
     "kind": "model",
     "note": "Model: 28,312 downloads, 86 likes."
    },
    {
     "name": "bioptimus/H0-mini",
     "url": "https://huggingface.co/bioptimus/H0-mini",
     "kind": "model",
     "note": "Model: 12,083 downloads, 43 likes."
    }
   ],
   "papers": [
    {
     "title": "Distilling foundation models for robust and efficient models in digital pathology",
     "url": "https://arxiv.org/abs/2501.16239",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "H-optimus-0",
     "url": "https://huggingface.co/bioptimus/H-optimus-0",
     "note": "Pathology foundation model"
    },
    {
     "name": "H-optimus-1",
     "url": "https://huggingface.co/bioptimus/H-optimus-1",
     "note": "Larger pathology foundation model"
    },
    {
     "name": "H0-mini",
     "url": "https://huggingface.co/bioptimus/H0-mini",
     "note": "Distilled pathology model"
    }
   ],
   "openness": "Selected models are available on Hugging Face; check each model card for its licence, which may restrict commercial use.",
   "key_people": [
    {
     "name": "Jean-Philippe Vert",
     "role": "Co-founder and CEO",
     "url": "https://www.bioptimus.com/about",
     "note": null
    },
    {
     "name": "Rodolphe Jenatton",
     "role": "Co-founder and CTO",
     "url": "https://www.bioptimus.com/about",
     "note": null
    },
    {
     "name": "Zelda Mariet",
     "role": "Co-founder and VP of Research",
     "url": "https://www.bioptimus.com/about",
     "note": null
    },
    {
     "name": "Felipe Llinares",
     "role": "Co-founder and VP of AI",
     "url": "https://www.bioptimus.com/about",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website.",
   "sources": [
    "https://www.bioptimus.com",
    "https://huggingface.co/bioptimus",
    "https://github.com/bioptimus"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "bioptimus",
   "mlat": 48.78042,
   "mlon": 2.45775,
   "logo": true
  },
  {
   "orig_id": "black-forest-labs",
   "name": "Black Forest Labs",
   "alias": null,
   "full_name": "Black Forest Labs",
   "type": "startup",
   "founded": 2024,
   "country": "Germany",
   "city": "Freiburg",
   "lat": 47.999,
   "lon": 7.8421,
   "website": "https://bfl.ai",
   "huggingface": "https://huggingface.co/black-forest-labs",
   "github": "https://github.com/black-forest-labs",
   "x": "https://x.com/bfl_ai",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/bflai",
   "contact_email": "info@blackforestlabs.ai",
   "summary": "Image and video generation lab behind the FLUX model family. It releases several open-weight variants alongside commercial API models.",
   "tagline": "A Freiburg lab founded by the creators of Stable Diffusion that builds the FLUX image models, with open-weight and API tiers.",
   "about": [
    "Black Forest Labs was founded in 2024 in Freiburg im Breisgau by Robin Rombach, Andreas Blattmann and Patrick Esser, former Stability AI researchers who had worked on latent diffusion at LMU Munich.",
    "FLUX.1 appeared in August 2024 and FLUX.1 Kontext in May 2025, followed by FLUX.2 in November 2025. FLUX.1 Schnell is Apache 2.0, while Dev variants are open weights under non-commercial licences and Pro models are API only.",
    "Early investors included Andreessen Horowitz, and a reported 450 million dollar financing followed in late 2025."
   ],
   "mission": "The lab aims to advance generative image and video models while offering part of its work as open weights.",
   "notice": "Black Forest Labs is a commercial company. Some FLUX models are open weights under non-commercial licences, while the strongest models are API-only.",
   "known_for": [
    "FLUX.1 image generation models",
    "FLUX.1 Kontext editing models",
    "Founders from the Stable Diffusion team"
   ],
   "focus": [
    "image generation",
    "vision",
    "diffusion"
   ],
   "contributions": [
    {
     "title": "FLUX.1 released",
     "note": "Rectified-flow transformer image models scaled to 12 billion parameters.",
     "year": "2024"
    },
    {
     "title": "FLUX.1 Kontext released",
     "note": "In-context image generation and editing models.",
     "year": "2025"
    },
    {
     "title": "FLUX.2 released",
     "note": "Models with Pro, Flex, Dev and Klein variants and an open VAE.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "flux",
     "url": "https://github.com/black-forest-labs/flux",
     "note": "Official inference repo for FLUX.1 models (26,063 stars)"
    },
    {
     "name": "flux2",
     "url": "https://github.com/black-forest-labs/flux2",
     "note": "Official inference repo for FLUX.2 models (2,879 stars)"
    },
    {
     "name": "Self-Flow",
     "url": "https://github.com/black-forest-labs/Self-Flow",
     "note": "[ICML'26] Code and website for Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis (795 stars)"
    },
    {
     "name": "flux-action",
     "url": "https://github.com/black-forest-labs/flux-action",
     "note": "FLUX 3 Action: open weights 7B world action model from Black Forest Labs. Train, fine-tune and run action pred (222 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "black-forest-labs/flux-3-image-klein-comfy",
     "url": "https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy",
     "kind": "model",
     "note": "Model: 1,115,161 downloads, 3355 likes."
    },
    {
     "name": "black-forest-labs/FLUX.2-klein-base-4B",
     "url": "https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B",
     "kind": "model",
     "note": "Model: 731,723 downloads, 197 likes."
    },
    {
     "name": "black-forest-labs/kontext-bench",
     "url": "https://huggingface.co/datasets/black-forest-labs/kontext-bench",
     "kind": "dataset",
     "note": "Dataset: 385 downloads, 68 likes."
    },
    {
     "name": "black-forest-labs/FLUX.1-dev",
     "url": "https://huggingface.co/spaces/black-forest-labs/FLUX.1-dev",
     "kind": "space",
     "note": "Space: 9584 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space",
     "url": "https://arxiv.org/abs/2506.15742",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "FLUX.1 [dev]",
     "url": "https://huggingface.co/black-forest-labs/FLUX.1-dev",
     "note": "Open-weight text-to-image model (non-commercial licence)"
    },
    {
     "name": "FLUX.1 [schnell]",
     "url": "https://huggingface.co/black-forest-labs/FLUX.1-schnell",
     "note": "Fast open-weight text-to-image model"
    },
    {
     "name": "flux",
     "url": "https://github.com/black-forest-labs/flux",
     "note": "Official inference code for FLUX.1"
    }
   ],
   "openness": "Some FLUX models (e.g. schnell) are Apache-2.0; others are open weights under non-commercial licences; inference code is Apache-2.0.",
   "key_people": [
    {
     "name": "Robin Rombach",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Black_Forest_Labs",
     "note": null
    },
    {
     "name": "Andreas Blattmann",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Black_Forest_Labs",
     "note": null
    },
    {
     "name": "Patrick Esser",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Black_Forest_Labs",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on the website.",
   "sources": [
    "https://bfl.ai",
    "https://huggingface.co/black-forest-labs",
    "https://github.com/black-forest-labs",
    "https://arxiv.org/abs/2506.15742"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "black-forest-labs",
   "mlat": 48.02065,
   "mlon": 7.89814,
   "logo": true
  },
  {
   "orig_id": "boltz",
   "name": "Boltz",
   "alias": null,
   "full_name": "Boltz",
   "type": "startup",
   "founded": 2025,
   "country": "United States",
   "city": "Cambridge",
   "lat": 42.3736,
   "lon": -71.1097,
   "website": "https://boltz.com",
   "huggingface": "https://huggingface.co/boltz-community",
   "github": "https://github.com/jwohlwend",
   "x": "https://x.com/boltz_bio",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Public-benefit company growing out of MIT research that develops open biomolecular foundation models (Boltz-1, Boltz-2) for structure and binding-affinity prediction.",
   "tagline": "A public benefit company growing out of MIT research that develops biomolecular models, with open Boltz-2 and BoltzGen alongside proprietary ones.",
   "about": [
    "Boltz PBC builds AI models and infrastructure for biomolecular design and drug discovery. Its platform, Boltz Lab, lets scientists design small molecules and biologics.",
    "Its open-source models include Boltz-2 and BoltzGen, while BoltzProt-1 and BoltzMol-1 are proprietary. The site reports more than 5 million downloads and over 1,000 companies using its tools, and announced an agreement with Bristol Myers Squibb in October 2026."
   ],
   "mission": "Boltz aims to make biomolecular modelling broadly available while building a commercial platform on top.",
   "notice": null,
   "known_for": [
    "Boltz-1 and Boltz-2 open biomolecular models",
    "BoltzGen design model",
    "Boltz Lab platform"
   ],
   "focus": [
    "biomolecular modelling",
    "structure prediction",
    "drug discovery"
   ],
   "contributions": [
    {
     "title": "Boltz-1 released",
     "note": "The first open model from the MIT group, released in late 2024.",
     "year": "2024"
    },
    {
     "title": "Boltz-2 released",
     "note": "A model that jointly predicts complex structure and binding affinity.",
     "year": "2025"
    },
    {
     "title": "Seed round and Pfizer collaboration",
     "note": "Boltz PBC launched with a 28 million dollar seed round alongside Boltz Lab and a Pfizer collaboration.",
     "year": "2026"
    },
    {
     "title": "Bristol Myers Squibb agreement",
     "note": "An agreement to deploy biomolecular AI models for drug discovery.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "boltz",
     "url": "https://github.com/jwohlwend/boltz",
     "note": "Official Boltz repository"
    },
    {
     "name": "minifold",
     "url": "https://github.com/jwohlwend/minifold",
     "note": "Fast protein structure prediction"
    }
   ],
   "hf_highlights": [
    {
     "name": "boltz-1",
     "url": "https://huggingface.co/boltz-community/boltz-1",
     "kind": "model",
     "note": "Boltz-1 weights"
    },
    {
     "name": "boltz-2",
     "url": "https://huggingface.co/boltz-community/boltz-2",
     "kind": "model",
     "note": "Boltz-2 weights"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "boltz",
     "url": "https://github.com/jwohlwend/boltz",
     "note": "Official Boltz repository"
    },
    {
     "name": "minifold",
     "url": "https://github.com/jwohlwend/minifold",
     "note": "Fast protein structure prediction"
    }
   ],
   "openness": "MIT-licensed code and weights.",
   "key_people": [
    {
     "name": "Gabriele Corso",
     "role": "Co-founder and CEO",
     "url": "https://gcorso.github.io/",
     "note": "Previously at MIT CSAIL."
    },
    {
     "name": "Jeremy Wohlwend",
     "role": "Co-founder and CTO",
     "url": "https://github.com/jwohlwend",
     "note": "Maintains the Boltz repository."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://boltz.com/",
    "https://github.com/jwohlwend/boltz",
    "https://huggingface.co/boltz-community"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "boltz",
   "mlat": 42.33824,
   "mlon": -71.1097,
   "logo": true
  },
  {
   "orig_id": "browser-use",
   "name": "Browser Use",
   "alias": null,
   "full_name": "Browser Use",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://browser-use.com",
   "huggingface": "https://huggingface.co/browser-use",
   "github": "https://github.com/browser-use",
   "x": "https://x.com/browser_use",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source library and company for AI browser agents, with a small open model.",
   "tagline": "An open-source library and company for AI browser agents, with hosted agents and stealth browsers.",
   "about": [
    "Browser Use began as an open-source Python library that lets language-model agents control a web browser, built by Magnus Müller and Gregor Žunič, who met during master's studies at ETH Zurich. It joined Y Combinator's Winter 2025 batch.",
    "The company raised a 17 million dollar seed round led by Felicis in March 2025. Today it offers hosted web agents, stealth cloud browsers and Browser Harness, a free open-source tool for coding agents, with enterprise options such as on-premises browsers and an EU cloud in Frankfurt."
   ],
   "mission": "The company aims to let AI agents use the web reliably.",
   "notice": null,
   "known_for": [
    "Open-source browser-use agent library",
    "Cloud and stealth browsers",
    "Browser Harness for coding agents"
   ],
   "focus": [
    "agents",
    "browser automation"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded by two ETH Zurich master's graduates and part of Y Combinator's Winter 2025 batch.",
     "year": "2024"
    },
    {
     "title": "Seed round",
     "note": "Raised a 17 million dollar seed round led by Felicis in March 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "browser-use",
     "url": "https://github.com/browser-use/browser-use",
     "note": "Python library for AI browser agents"
    },
    {
     "name": "browser-harness",
     "url": "https://github.com/browser-use/browser-harness",
     "note": "Real Chrome automation for coding agents"
    }
   ],
   "hf_highlights": [
    {
     "name": "bu-30b-a3b-preview",
     "url": "https://huggingface.co/browser-use/bu-30b-a3b-preview",
     "kind": "model",
     "note": "Open browser agent model"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "browser-use",
     "url": "https://github.com/browser-use/browser-use",
     "note": "Python library for AI browser agents"
    },
    {
     "name": "browser-harness",
     "url": "https://github.com/browser-use/browser-harness",
     "note": "Real Chrome automation for coding agents"
    }
   ],
   "openness": "MIT-licensed library and model.",
   "key_people": [
    {
     "name": "Magnus Müller",
     "role": "Co-founder",
     "url": "https://github.com/MagMueller",
     "note": null
    },
    {
     "name": "Gregor Žunič",
     "role": "Co-founder",
     "url": "https://github.com/gregpr07",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://browser-use.com",
    "https://github.com/browser-use",
    "https://huggingface.co/browser-use"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "browser-use",
   "mlat": 37.67937,
   "mlon": -122.32752,
   "logo": true
  },
  {
   "orig_id": "bsc-langtech",
   "name": "BSC Language Technologies Lab",
   "alias": "Salamandra, ALIA",
   "full_name": "BSC Language Technologies Lab (Salamandra, ALIA)",
   "type": "institute",
   "founded": null,
   "country": "Spain",
   "city": "Barcelona",
   "lat": 41.3874,
   "lon": 2.1686,
   "website": "https://langtech.bsc.es",
   "huggingface": "https://huggingface.co/BSC-LT",
   "github": "https://github.com/langtech-bsc",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Language technologies group at the Barcelona Supercomputing Center that trains the Salamandra and ALIA open models for Spanish, Catalan and other European languages.",
   "tagline": "The Barcelona Supercomputing Center language technologies group that trains the open Salamandra and ALIA models for Iberian and European languages.",
   "about": [
    "The Language Technologies unit at the Barcelona Supercomputing Center, led by Marta Villegas, develops resources and models for Spanish, Catalan and other European languages.",
    "ALIA, a public, open and multilingual AI infrastructure coordinated by BSC, was announced in January 2025. Its Salamandra models, in 2B, 7B and 40B sizes, are published on Hugging Face under Apache 2.0 and were trained on MareNostrum 5."
   ],
   "mission": "The group aims to build open language technology for Spain's languages with public funding.",
   "notice": null,
   "known_for": [
    "Salamandra open models",
    "ALIA public AI infrastructure",
    "MareNostrum 5 training"
   ],
   "focus": [
    "language models",
    "Spanish",
    "Catalan",
    "speech"
   ],
   "contributions": [
    {
     "title": "Salamandra technical report",
     "note": "Report on open 2B, 7B and 40B models published February 2025.",
     "year": "2025"
    },
    {
     "title": "ALIA announced",
     "note": "Spain announced the public AI project coordinated by BSC.",
     "year": "2025"
    },
    {
     "title": "Salamandra released",
     "note": "Open 2B, 7B and 40B models on Hugging Face.",
     "year": "2025"
    },
    {
     "title": "ALIA-40b",
     "note": "A 40B model pretrained from scratch on 9.37 trillion tokens, per its model card.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "salamandra",
     "url": "https://github.com/langtech-bsc/salamandra",
     "note": "Open-source repository (33 stars)"
    },
    {
     "name": "Wikiextractor-V2",
     "url": "https://github.com/langtech-bsc/Wikiextractor-V2",
     "note": "Enhaced version of Wikiextrator: A wikipedia dumps extractor (31 stars)"
    },
    {
     "name": "alia",
     "url": "https://github.com/langtech-bsc/alia",
     "note": "Open-source repository (28 stars)"
    },
    {
     "name": "AnonymizationPipeline",
     "url": "https://github.com/langtech-bsc/AnonymizationPipeline",
     "note": "Anonymization Pipeline for injesting data from outside of BSC that contains GDPR protected data. (19 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "BSC-LT/salamandra-7b-instruct",
     "url": "https://huggingface.co/BSC-LT/salamandra-7b-instruct",
     "kind": "model",
     "note": "Model: 33,786 downloads, 85 likes."
    },
    {
     "name": "BSC-LT/MrBERT-es",
     "url": "https://huggingface.co/BSC-LT/MrBERT-es",
     "kind": "model",
     "note": "Model: 11,093 downloads, 12 likes."
    },
    {
     "name": "BSC-LT/EsBBQ",
     "url": "https://huggingface.co/datasets/BSC-LT/EsBBQ",
     "kind": "dataset",
     "note": "Dataset: 1,481 downloads, 0 likes."
    },
    {
     "name": "BSC-LT/SalamandraTA-7B-Demo",
     "url": "https://huggingface.co/spaces/BSC-LT/SalamandraTA-7B-Demo",
     "kind": "space",
     "note": "Space: 6 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Salamandra Technical Report",
     "url": "https://arxiv.org/abs/2502.08489",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "ALIA-40b",
     "url": "https://huggingface.co/BSC-LT/ALIA-40b",
     "note": "Open 40B model for Iberian languages"
    },
    {
     "name": "Salamandra-7b-instruct",
     "url": "https://huggingface.co/BSC-LT/salamandra-7b-instruct",
     "note": "Multilingual instruction model"
    },
    {
     "name": "salamandra",
     "url": "https://github.com/langtech-bsc/salamandra",
     "note": "Salamandra project repository"
    }
   ],
   "openness": "Models are published openly on Hugging Face; the Salamandra code repositories are Apache-2.0.",
   "key_people": [
    {
     "name": "Marta Villegas",
     "role": "Head of the Language Technologies Unit",
     "url": "https://www.xataka.com/robotica-e-ia/asi-alia-modelo-ia-espanola-que-puede-quiere-competir-chatgpt",
     "note": null
    },
    {
     "name": "Aitor Gonzalez-Agirre",
     "role": "Lead author, Salamandra Technical Report",
     "url": "https://arxiv.org/abs/2502.08489",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the BSC website.",
   "sources": [
    "https://langtech.bsc.es",
    "https://huggingface.co/BSC-LT",
    "https://github.com/langtech-bsc",
    "https://arxiv.org/abs/2502.08489"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "bsc-language-technologies-lab",
   "mlat": 41.3374,
   "mlon": 2.1686,
   "logo": true
  },
  {
   "orig_id": "buet-csebuetnlp",
   "name": "BUET CSE NLP Group",
   "alias": null,
   "full_name": "BUET CSE NLP Group",
   "type": "academic",
   "founded": null,
   "country": "Bangladesh",
   "city": "Dhaka",
   "lat": 23.81,
   "lon": 90.41,
   "website": "https://csebuetnlp.github.io/",
   "huggingface": "https://huggingface.co/csebuetnlp",
   "github": "https://github.com/csebuetnlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "University research group on low-resource Bangla NLP, releasing BanglaBERT and BanglaT5.",
   "tagline": "A BUET research group on low-resource Bangla NLP that releases BanglaBERT, BanglaT5 and the XL-Sum summarisation dataset.",
   "about": [
    "The CSE NLP group at Bangladesh University of Engineering and Technology works on machine translation, multilingual NLP and language understanding. Its faculty include Rifat Shahriyar, M. Sohel Rahman and Anindya Iqbal.",
    "It open-sourced BanglaBERT in 2021 and BanglaT5 in 2022, and published XL-Sum, a 44-language summarisation dataset, at ACL 2021 Findings."
   ],
   "mission": "The group aims to provide open models and benchmarks for Bangla.",
   "notice": null,
   "known_for": [
    "BanglaBERT",
    "BanglaT5",
    "XL-Sum dataset"
   ],
   "focus": [
    "Bangla",
    "NLP",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Bengali-English MT datasets",
     "note": "Aligner ensembling and new datasets for Bengali-English translation were published.",
     "year": "2020"
    },
    {
     "title": "XL-Sum published",
     "note": "A summarisation dataset in 44 languages, at Findings of ACL 2021.",
     "year": "2021"
    },
    {
     "title": "BanglaBERT open-sourced",
     "note": "A Bangla language model, later published in Findings of NAACL 2022.",
     "year": "2021"
    },
    {
     "title": "BanglaT5 released",
     "note": "A Bangla sequence-to-sequence model.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "xl-sum",
     "url": "https://github.com/csebuetnlp/xl-sum",
     "note": "XL-Sum code, data, models."
    },
    {
     "name": "banglabert",
     "url": "https://github.com/csebuetnlp/banglabert",
     "note": "BanglaBERT release."
    },
    {
     "name": "banglanmt",
     "url": "https://github.com/csebuetnlp/banglanmt",
     "note": "Bangla NMT."
    },
    {
     "name": "BanglaNLG",
     "url": "https://github.com/csebuetnlp/BanglaNLG",
     "note": "BanglaT5 release."
    }
   ],
   "hf_highlights": [
    {
     "name": "csebuetnlp/banglabert",
     "url": "https://huggingface.co/csebuetnlp/banglabert",
     "kind": "model",
     "note": "BanglaBERT."
    },
    {
     "name": "csebuetnlp/mT5_multilingual_XLSum",
     "url": "https://huggingface.co/csebuetnlp/mT5_multilingual_XLSum",
     "kind": "model",
     "note": "Multilingual summarizer."
    },
    {
     "name": "csebuetnlp/xlsum",
     "url": "https://huggingface.co/datasets/csebuetnlp/xlsum",
     "kind": "dataset",
     "note": "XL-Sum."
    },
    {
     "name": "csebuetnlp/CrossSum",
     "url": "https://huggingface.co/datasets/csebuetnlp/CrossSum",
     "kind": "dataset",
     "note": "Cross-lingual summarization."
    }
   ],
   "papers": [
    {
     "title": "XL-Sum: Large-Scale Multilingual Abstractive Summarization for 44 Languages",
     "url": "https://arxiv.org/abs/2106.13822",
     "year": "2021"
    },
    {
     "title": "BanglaBERT: Language Model Pretraining and Benchmarks for Low-Resource Language Understanding Evaluation in Bangla",
     "url": "https://arxiv.org/abs/2101.00204",
     "year": "2021"
    },
    {
     "title": "BanglaNLG and BanglaT5: Benchmarks and Resources for Evaluating Low-Resource Natural Language Generation in Bangla",
     "url": "https://arxiv.org/abs/2205.11081",
     "year": "2022"
    },
    {
     "title": "Not Low-Resource Anymore: Aligner Ensembling, Batch Filtering, and New Datasets for Bengali-English Machine Translation",
     "url": "https://arxiv.org/abs/2009.09359",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "banglabert",
     "url": "https://huggingface.co/csebuetnlp/banglabert",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "banglat5_nmt_bn_en",
     "url": "https://huggingface.co/csebuetnlp/banglat5_nmt_bn_en",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open models and datasets.",
   "key_people": [
    {
     "name": "Rifat Shahriyar",
     "role": "Professor, BUET",
     "url": "https://rifatshahriyar.github.io/",
     "note": null
    },
    {
     "name": "M. Sohel Rahman",
     "role": "Professor, BUET",
     "url": "https://csebuetnlp.github.io/",
     "note": null
    },
    {
     "name": "Anindya Iqbal",
     "role": "Professor, BUET",
     "url": "https://csebuetnlp.github.io/",
     "note": null
    },
    {
     "name": "Tahmid Hasan",
     "role": "Researcher; lead author of XL-Sum",
     "url": "https://arxiv.org/abs/2106.13822",
     "note": null
    },
    {
     "name": "Abhik Bhattacharjee",
     "role": "Research assistant; lead author of BanglaBERT and BanglaT5",
     "url": "https://arxiv.org/abs/2101.00204",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The group site csebuetnlp.github.io links its GitHub and papers; GitHub issues are the open route.",
   "sources": [
    "https://huggingface.co/csebuetnlp",
    "https://csebuetnlp.github.io/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "buet-cse-nlp-group",
   "logo": true
  },
  {
   "orig_id": "but-fit",
   "name": "BUT Speech@FIT",
   "alias": null,
   "full_name": "BUT Speech@FIT",
   "type": "academic",
   "founded": null,
   "country": "Czechia",
   "city": "Brno",
   "lat": 49.1951,
   "lon": 16.6068,
   "website": "https://speech.fit.vut.cz",
   "huggingface": "https://huggingface.co/BUT-FIT",
   "github": "https://github.com/BUTSpeechFIT",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "speech-web@fit.vut.cz",
   "summary": "Speech and language research group at Brno University of Technology, open-sourcing speaker diarization toolkits and Czech language models.",
   "tagline": "Brno University of Technology's speech group, author of the VBx and DiariZen diarization toolkits and Czech language models.",
   "about": [
    "BUT Speech@FIT is the speech and language research group at the Faculty of Information Technology, Brno University of Technology. Its GitHub organization dates from 2019.",
    "The group released VBx, a Bayesian HMM clustering approach to speaker diarization described by Landini, Profant, Diez and Burget, and DiariZen, a pipeline built on pyannote with WavLM-based models. Jan Černocký and Lukáš Burget are co-authors on its diarization system papers.",
    "On Hugging Face the group published csmpt7b, a Czech-focused 7B model, in 2024. Its tools serve speech researchers and diarization challenge participants."
   ],
   "mission": "The group advances speech processing and shares open recipes and models, especially for diarization and Czech language technology.",
   "notice": null,
   "known_for": [
    "VBx diarization",
    "DiariZen toolkit",
    "Czech LLM csmpt7b",
    "Speech technology challenge systems"
   ],
   "focus": [
    "speech processing",
    "speaker diarization",
    "Czech LLMs"
   ],
   "contributions": [
    {
     "title": "VBx",
     "note": "Released an open implementation of a widely used diarization recipe.",
     "year": "2020"
    },
    {
     "title": "csmpt7b",
     "note": "Published a Czech-focused 7B model on Hugging Face.",
     "year": "2024"
    },
    {
     "title": "DiariZen",
     "note": "Released a diarization toolkit with pre-trained WavLM-based models.",
     "year": "2024"
    },
    {
     "title": "MLC-SLM diarization",
     "note": "A 2025 paper reported DiariZen outperforming the baseline in multilingual tests.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "DiariZen",
     "url": "https://github.com/BUTSpeechFIT/DiariZen",
     "note": "A toolkit for speaker diarization.  (569 stars)"
    },
    {
     "name": "VBx",
     "url": "https://github.com/BUTSpeechFIT/VBx",
     "note": "Variational Bayes HMM over x-vectors diarization (291 stars)"
    },
    {
     "name": "speakerbeam",
     "url": "https://github.com/BUTSpeechFIT/speakerbeam",
     "note": "Open-source repository. (158 stars)"
    },
    {
     "name": "TS-ASR-Whisper",
     "url": "https://github.com/BUTSpeechFIT/TS-ASR-Whisper",
     "note": "Open-source repository. (128 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "BUT-FIT/diarizen-wavlm-large-s80-md",
     "url": "https://huggingface.co/BUT-FIT/diarizen-wavlm-large-s80-md",
     "kind": "model",
     "note": "Model with 56 likes on the Hub."
    },
    {
     "name": "BUT-FIT/diarizen-wavlm-large-s80-md-v2",
     "url": "https://huggingface.co/BUT-FIT/diarizen-wavlm-large-s80-md-v2",
     "kind": "model",
     "note": "Model with 28 likes on the Hub."
    },
    {
     "name": "BUT-FIT/BUT-LCC",
     "url": "https://huggingface.co/datasets/BUT-FIT/BUT-LCC",
     "kind": "dataset",
     "note": "Dataset with 9 likes on the Hub."
    },
    {
     "name": "BUT-FIT/EMMA_leaderboard",
     "url": "https://huggingface.co/spaces/BUT-FIT/EMMA_leaderboard",
     "kind": "space",
     "note": "Space with 14 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Leveraging Self-Supervised Learning for Speaker Diarization",
     "url": "https://arxiv.org/abs/2409.09408",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "DiariZen",
     "url": "https://github.com/BUTSpeechFIT/DiariZen",
     "note": "Speaker diarization toolkit"
    },
    {
     "name": "VBx",
     "url": "https://github.com/BUTSpeechFIT/VBx",
     "note": "Variational Bayes x-vector diarization"
    },
    {
     "name": "csmpt7b",
     "url": "https://huggingface.co/BUT-FIT/csmpt7b",
     "note": "Czech 7B language model (Apache-2.0)"
    },
    {
     "name": "TS-ASR-Whisper",
     "url": "https://github.com/BUTSpeechFIT/TS-ASR-Whisper",
     "note": "Target-speaker ASR with Whisper"
    }
   ],
   "openness": "Toolkits and models are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Lukáš Burget",
     "role": "Researcher and diarization co-author",
     "url": "https://speech.fit.vut.cz",
     "note": null
    },
    {
     "name": "Jan Černocký",
     "role": "Researcher and diarization co-author",
     "url": "https://speech.fit.vut.cz",
     "note": null
    },
    {
     "name": "Federico Landini",
     "role": "VBx lead author",
     "url": "https://arxiv.org/abs/2012.14952",
     "note": null
    },
    {
     "name": "Martin Kocour",
     "role": "Member of the BUTSpeechFIT GitHub organization",
     "url": "https://www.vutbr.cz/en/people/martin-kocour-165504",
     "note": null
    },
    {
     "name": "Lucas Ondel",
     "role": "Member of the BUTSpeechFIT GitHub organization",
     "url": "https://lucasondel.github.io/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://speech.fit.vut.cz",
    "https://huggingface.co/BUT-FIT",
    "https://github.com/BUTSpeechFIT",
    "https://arxiv.org/abs/2409.09408"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "but-speech-fit",
   "logo": true
  },
  {
   "orig_id": "canopy-labs",
   "name": "Canopy Labs",
   "alias": null,
   "full_name": "Canopy Labs",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://canopylabs.ai",
   "huggingface": "https://huggingface.co/canopylabs",
   "github": "https://github.com/canopyai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Speech AI lab behind Orpheus, an Apache-2.0 family of Llama-based text-to-speech models.",
   "tagline": "A speech AI lab behind Orpheus, a family of Llama-based text-to-speech models released in March 2025 for open use.",
   "about": [
    "Canopy Labs released Orpheus in March 2025, a text-to-speech system built on a Llama 3B backbone, with models published on GitHub and Hugging Face.",
    "The family is available in sizes from 150M to 3B parameters. Sources disagree on the exact weight licence, so users should check the repository. The lab has not published a public team page."
   ],
   "mission": "The lab develops expressive, low-latency speech models.",
   "notice": null,
   "known_for": [
    "Orpheus TTS",
    "Emotion-tagged speech synthesis"
   ],
   "focus": [
    "text-to-speech",
    "speech",
    "open models"
   ],
   "contributions": [
    {
     "title": "Orpheus released",
     "note": "A Llama-based TTS family released as open source.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "Orpheus-TTS",
     "url": "https://github.com/canopyai/Orpheus-TTS",
     "note": "Towards human-sounding speech"
    }
   ],
   "hf_highlights": [
    {
     "name": "orpheus-3b-0.1-ft",
     "url": "https://huggingface.co/canopylabs/orpheus-3b-0.1-ft",
     "kind": "model",
     "note": "Fine-tuned Orpheus 3B"
    },
    {
     "name": "orpheus-3b-0.1-pretrained",
     "url": "https://huggingface.co/canopylabs/orpheus-3b-0.1-pretrained",
     "kind": "model",
     "note": "Pretrained Orpheus 3B"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Orpheus-TTS",
     "url": "https://github.com/canopyai/Orpheus-TTS",
     "note": "Towards human-sounding speech"
    }
   ],
   "openness": "Apache-2.0 Orpheus code and weights.",
   "key_people": [],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://canopylabs.ai",
    "https://github.com/canopyai/Orpheus-TTS",
    "https://huggingface.co/canopylabs"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "canopy-labs",
   "mlat": 37.69721,
   "mlon": -122.30369,
   "logo": true
  },
  {
   "orig_id": "cardiffnlp",
   "name": "Cardiff NLP",
   "alias": null,
   "full_name": "Cardiff NLP",
   "type": "academic",
   "founded": 2020,
   "country": "United Kingdom",
   "city": "Cardiff",
   "lat": 51.4816,
   "lon": -3.1791,
   "website": "https://www.cardiffnlp.org",
   "huggingface": "https://huggingface.co/cardiffnlp",
   "github": "https://github.com/cardiffnlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Cardiff NLP is the NLP group at Cardiff University, known for TweetEval, TweetNLP and the Twitter-RoBERTa models used widely for social media analysis.",
   "tagline": "Cardiff University's NLP group, known for TweetEval, TweetNLP and the Twitter-RoBERTa models for social media analysis.",
   "about": [
    "Cardiff NLP is a research group at Cardiff University whose GitHub organization was created in October 2020. Jose Camacho-Collados is a prominent member and an author on its flagship benchmark.",
    "TweetEval, presented in Findings of EMNLP 2020, standardises evaluation across tweet classification tasks. The group also released XLM-T and TimeLMs, and the TweetNLP Python platform in 2022, with Twitter-RoBERTa models trained on tens of millions of tweets.",
    "Its models are widely downloaded from Hugging Face and serve researchers and practitioners analysing social media."
   ],
   "mission": "The group aims to build language models and evaluation resources that handle the informal, fast-changing language of social media, and to release them openly.",
   "notice": null,
   "known_for": [
    "TweetEval benchmark",
    "TweetNLP platform",
    "Twitter-RoBERTa models",
    "TimeLMs diachronic models"
   ],
   "focus": [
    "social media NLP"
   ],
   "contributions": [
    {
     "title": "TweetEval",
     "note": "Created a unified benchmark for tweet classification (arXiv 2010.12421).",
     "year": "2020"
    },
    {
     "title": "XLM-T",
     "note": "Released multilingual Twitter language models (arXiv 2104.12250).",
     "year": "2021"
    },
    {
     "title": "TimeLMs",
     "note": "Released diachronic language models (arXiv 2202.03829).",
     "year": "2022"
    },
    {
     "title": "TweetNLP",
     "note": "Released a Python platform for social media NLP built on Twitter-trained models.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "tweeteval",
     "url": "https://github.com/cardiffnlp/tweeteval",
     "note": "Repository for TweetEval (406 stars)"
    },
    {
     "name": "tweetnlp",
     "url": "https://github.com/cardiffnlp/tweetnlp",
     "note": "TweetNLP for all the NLP enthusiasts working on Twitter! The Python library tweetnlp provides a collection of  (395 stars)"
    },
    {
     "name": "xlm-t",
     "url": "https://github.com/cardiffnlp/xlm-t",
     "note": "Repository for XLM-T, a framework for evaluating multilingual language models on Twitter data (167 stars)"
    },
    {
     "name": "timelms",
     "url": "https://github.com/cardiffnlp/timelms",
     "note": "TimeLMs: Diachronic Language Models from Twitter (114 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "cardiffnlp/twitter-roberta-base-sentiment-latest",
     "url": "https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment-latest",
     "kind": "model",
     "note": "Model with 858 likes on the Hub."
    },
    {
     "name": "cardiffnlp/twitter-roberta-base-sentiment",
     "url": "https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment",
     "kind": "model",
     "note": "Model with 343 likes on the Hub."
    },
    {
     "name": "cardiffnlp/tweet_eval",
     "url": "https://huggingface.co/datasets/cardiffnlp/tweet_eval",
     "kind": "dataset",
     "note": "Dataset with 164 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification",
     "url": "https://arxiv.org/abs/2010.12421",
     "year": "2020"
    },
    {
     "title": "TimeLMs: Diachronic Language Models from Twitter",
     "url": "https://arxiv.org/abs/2202.03829",
     "year": "2022"
    },
    {
     "title": "XLM-T: Multilingual Language Models in Twitter for Sentiment Analysis and Beyond",
     "url": "https://arxiv.org/abs/2104.12250",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "twitter-roberta-base-sentiment-latest",
     "url": "https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment-latest",
     "note": "Sentiment model with millions of downloads"
    },
    {
     "name": "tweet_eval",
     "url": "https://huggingface.co/datasets/cardiffnlp/tweet_eval",
     "note": "Unified benchmark for tweet classification"
    },
    {
     "name": "TweetNLP",
     "url": "https://github.com/cardiffnlp/tweetnlp",
     "note": "Python library for Twitter NLP"
    },
    {
     "name": "TimeLMs",
     "url": "https://github.com/cardiffnlp/timelms",
     "note": "Diachronic language models from Twitter"
    }
   ],
   "openness": "Models, datasets and libraries are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Jose Camacho-Collados",
     "role": "Professor, Cardiff University; TweetEval author",
     "url": "http://josecamachocollados.com/",
     "note": null
    },
    {
     "name": "Luis Espinosa Anke",
     "role": "Academic staff, Cardiff University",
     "url": "https://cardiffnlp.github.io/",
     "note": null
    },
    {
     "name": "Steven Schockaert",
     "role": "Professor, Cardiff University",
     "url": "https://cardiffnlp.github.io/",
     "note": null
    },
    {
     "name": "Fernando Alva-Manchego",
     "role": "Member of the cardiffnlp GitHub organization",
     "url": "https://feralvam.github.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.cardiffnlp.org",
    "https://huggingface.co/cardiffnlp",
    "https://github.com/cardiffnlp",
    "https://arxiv.org/abs/2010.12421"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "cardiff-nlp",
   "logo": true
  },
  {
   "orig_id": "cartesia",
   "name": "Cartesia",
   "alias": null,
   "full_name": "Cartesia",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://cartesia.ai",
   "huggingface": "https://huggingface.co/cartesia-ai",
   "github": "https://github.com/cartesia-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Voice AI company that also publishes open state-space language models such as Rene and Llamba.",
   "tagline": "A San Francisco voice AI company founded by the Stanford authors of Mamba-style state-space models, which also releases open Rene and Llamba models.",
   "about": [
    "Cartesia was founded in 2023 by Karan Goel, Albert Gu, Arjun Desai and Brandon Yang, researchers who worked on state space models in Chris Re's group at Stanford. Goel serves as chief executive and Gu, a lead author of Mamba and a professor at Carnegie Mellon University, serves as chief scientist.",
    "The company builds real-time voice systems on state-space architectures, which process sequences with a fixed-size state rather than attention over the whole context. It raised a $22 million seed round led by Index Ventures, bringing its total to $27 million, and a $64 million Series A led by Kleiner Perkins in March 2025.",
    "Alongside its commercial speech products, Cartesia publishes open research artifacts on Hugging Face and GitHub. These include the Rene language model and the Llamba family of distilled recurrent models, described in a February 2025 paper that converts Llama models into recurrent form with a small fraction of the usual training data."
   ],
   "mission": "Cartesia aims to build real-time, on-device voice AI on state-space model architectures, and shares part of its model research openly.",
   "notice": null,
   "known_for": [
    "State-space model research from the authors of Mamba",
    "Open Rene and Llamba models",
    "Low-latency speech synthesis",
    "Distilling Transformers into recurrent models"
   ],
   "focus": [
    "state-space models",
    "speech"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Goel, Gu, Desai and Yang launched Cartesia to commercialize state-space model research.",
     "year": "2023"
    },
    {
     "title": "Seed round",
     "note": "A $22 million round led by Index Ventures brought total funding to $27 million.",
     "year": "2024"
    },
    {
     "title": "Llamba models",
     "note": "Distilled recurrent versions of Llama released with a paper and open weights.",
     "year": "2025"
    },
    {
     "title": "Series A",
     "note": "A $64 million round led by Kleiner Perkins, reported by Fortune.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "edge",
     "url": "https://github.com/cartesia-ai/edge",
     "note": "On-device intelligence."
    },
    {
     "name": "cartesia-python",
     "url": "https://github.com/cartesia-ai/cartesia-python",
     "note": "Python client."
    },
    {
     "name": "line",
     "url": "https://github.com/cartesia-ai/line",
     "note": "Voice agent SDK."
    }
   ],
   "hf_highlights": [
    {
     "name": "Llamba-1B",
     "url": "https://huggingface.co/cartesia-ai/Llamba-1B",
     "kind": "model",
     "note": "Llamba 1B."
    },
    {
     "name": "Rene-v0.1-1.3b-pytorch",
     "url": "https://huggingface.co/cartesia-ai/Rene-v0.1-1.3b-pytorch",
     "kind": "model",
     "note": "Rene."
    },
    {
     "name": "Llamba-8B",
     "url": "https://huggingface.co/cartesia-ai/Llamba-8B",
     "kind": "model",
     "note": "Llamba 8B."
    }
   ],
   "papers": [
    {
     "title": "Llamba: Scaling Distilled Recurrent Models for Efficient Language Processing",
     "url": "https://arxiv.org/abs/2502.14458",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Rene-v0.1-1.3b",
     "url": "https://huggingface.co/cartesia-ai/Rene-v0.1-1.3b-pytorch",
     "note": "SSM LM"
    }
   ],
   "openness": "Some open weights.",
   "key_people": [
    {
     "name": "Karan Goel",
     "role": "Co-founder and chief executive",
     "url": "https://krandiash.github.io",
     "note": "Stanford PhD under Chris Re; works on sequence modeling."
    },
    {
     "name": "Albert Gu",
     "role": "Co-founder and chief scientist",
     "url": "https://goombalab.github.io",
     "note": "Professor at Carnegie Mellon University and lead author of Mamba."
    },
    {
     "name": "Chris Re",
     "role": "Stanford professor, advisor to the founders",
     "url": "https://cs.stanford.edu/people/chrismre/",
     "note": "Press reports list him on the Cartesia team."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://cartesia.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co/cartesia-ai",
    "https://github.com/cartesia-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "cartesia",
   "mlat": 37.71869,
   "mlon": -122.28527,
   "logo": true
  },
  {
   "orig_id": "center-for-ai-safety",
   "name": "Center for AI Safety",
   "alias": null,
   "full_name": "Center for AI Safety",
   "type": "nonprofit",
   "founded": 2022,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://safe.ai",
   "huggingface": "https://huggingface.co/cais",
   "github": "https://github.com/centerforaisafety",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "contact@safe.ai",
   "summary": "Nonprofit AI safety research organisation that releases open benchmarks such as MMLU-related sets, HarmBench, WMDP and Humanity's Last Exam.",
   "tagline": "A San Francisco nonprofit reducing societal-scale AI risk through research, field-building and open benchmarks such as HarmBench and WMDP.",
   "about": [
    "The Center for AI Safety was started in 2022 and is led by Dan Hendrycks as executive and research director, with Oliver Zhang as managing director. Its stated purpose is to reduce societal-scale risks from artificial intelligence, and it organizes its staff into research, cloud and DevOps, projects and operations teams.",
    "CAIS has operated a compute cluster since March 2023 that gives university safety researchers access to large models, and it has run machine learning safety workshops, a Trojan detection competition and a philosophy fellowship. In May 2023 it published the Statement on AI Risk, signed by more than 600 researchers and public figures.",
    "Its research output includes widely used open evaluations. HarmBench standardizes tests of automated red teaming and refusal, WMDP measures hazardous knowledge and evaluates unlearning methods, and Humanity's Last Exam collects expert-written questions to measure frontier capability."
   ],
   "mission": "CAIS works to reduce societal-scale risks from AI through technical research, field-building and advocacy.",
   "notice": null,
   "known_for": [
    "Statement on AI Risk",
    "HarmBench red-teaming benchmark",
    "WMDP and Humanity's Last Exam benchmarks",
    "Compute cluster for safety researchers"
   ],
   "focus": [
    "safety",
    "evaluation",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Center founded",
     "note": "Established as a nonprofit focused on societal-scale AI risk.",
     "year": "2022"
    },
    {
     "title": "Compute cluster",
     "note": "Began providing GPU access to university safety researchers in March 2023.",
     "year": "2023"
    },
    {
     "title": "Statement on AI Risk",
     "note": "Signed by more than 600 AI experts and public figures.",
     "year": "2023"
    },
    {
     "title": "HarmBench",
     "note": "Open framework for evaluating automated red teaming and robust refusal.",
     "year": "2024"
    },
    {
     "title": "WMDP benchmark",
     "note": "Open benchmark for hazardous knowledge and unlearning, with dozens of co-authors across institutions.",
     "year": "2024"
    },
    {
     "title": "Humanity's Last Exam",
     "note": "Multimodal expert-written benchmark built with Scale AI and over a thousand contributors.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "hle",
     "url": "https://github.com/centerforaisafety/hle",
     "note": "Humanity's Last Exam."
    },
    {
     "name": "HarmBench",
     "url": "https://github.com/centerforaisafety/HarmBench",
     "note": "Red-teaming framework."
    },
    {
     "name": "wmdp",
     "url": "https://github.com/centerforaisafety/wmdp",
     "note": "Hazardous knowledge proxy."
    },
    {
     "name": "emergent-values",
     "url": "https://github.com/centerforaisafety/emergent-values",
     "note": "Utility Engineering code."
    }
   ],
   "hf_highlights": [
    {
     "name": "mmlu",
     "url": "https://huggingface.co/datasets/cais/mmlu",
     "kind": "dataset",
     "note": "MMLU."
    },
    {
     "name": "hle",
     "url": "https://huggingface.co/datasets/cais/hle",
     "kind": "dataset",
     "note": "HLE."
    },
    {
     "name": "wmdp",
     "url": "https://huggingface.co/datasets/cais/wmdp",
     "kind": "dataset",
     "note": "WMDP."
    },
    {
     "name": "HarmBench-Llama-2-13b-cls",
     "url": "https://huggingface.co/cais/HarmBench-Llama-2-13b-cls",
     "kind": "model",
     "note": "HarmBench classifier."
    }
   ],
   "papers": [
    {
     "title": "Humanity's Last Exam",
     "url": "https://arxiv.org/abs/2501.14249",
     "year": "2025"
    },
    {
     "title": "The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning",
     "url": "https://arxiv.org/abs/2403.03218",
     "year": "2024"
    },
    {
     "title": "HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal",
     "url": "https://arxiv.org/abs/2402.04249",
     "year": "2024"
    },
    {
     "title": "Measuring Massive Multitask Language Understanding",
     "url": "https://arxiv.org/abs/2009.03300",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "HLE",
     "url": "https://github.com/centerforaisafety/hle",
     "note": "Benchmark"
    },
    {
     "name": "HarmBench",
     "url": "https://github.com/centerforaisafety/HarmBench",
     "note": "Red-teaming"
    }
   ],
   "openness": "Open benchmarks and code.",
   "key_people": [
    {
     "name": "Dan Hendrycks",
     "role": "Executive and research director",
     "url": "https://danhendrycks.com",
     "note": "Creator of MMLU and the GELU activation."
    },
    {
     "name": "Oliver Zhang",
     "role": "Managing director",
     "url": "https://safe.ai/about",
     "note": "Listed on the CAIS leadership page."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://safe.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://safe.ai",
    "https://huggingface.co/cais"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "center-for-ai-safety",
   "mlat": 37.7428,
   "mlon": -122.27312,
   "logo": true
  },
  {
   "orig_id": "chcaa",
   "name": "Center for Humanities Computing",
   "alias": "Aarhus",
   "full_name": "Center for Humanities Computing (Aarhus)",
   "type": "academic",
   "founded": null,
   "country": "Denmark",
   "city": "Aarhus",
   "lat": 56.1629,
   "lon": 10.2039,
   "website": "https://chc.au.dk",
   "huggingface": "https://huggingface.co/chcaa",
   "github": "https://github.com/centre-for-humanities-computing",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Aarhus University centre applying NLP to the humanities, authors of DaCy and the Scandinavian Embedding Benchmarks.",
   "tagline": "Aarhus University centre applying NLP to the humanities, behind DaCy, odyCy and the Scandinavian Embedding Benchmark.",
   "about": [
    "The Center for Humanities Computing at Aarhus University was set up by the faculty with Kristoffer Nielbo as its first director, and Nielbo later became the first Danish Professor of Humanities Computing. It runs computational infrastructure for humanities researchers.",
    "Its NLP work includes DaCy, a unified Danish pipeline built on spaCy by Kenneth Enevoldsen, Lasse Hansen and Nielbo, odyCy for Ancient Greek, and the Scandinavian Embedding Benchmark co-created with partners.",
    "The centre's GitHub organization has about 188 repositories and its models serve Nordic NLP researchers and humanities scholars."
   ],
   "mission": "The centre seeks to make computational methods available across the humanities and to provide open Danish and Scandinavian language tools.",
   "notice": null,
   "known_for": [
    "DaCy Danish NLP pipeline",
    "Scandinavian Embedding Benchmark",
    "odyCy Ancient Greek models",
    "Humanities computing infrastructure"
   ],
   "focus": [
    "Danish NLP",
    "embeddings",
    "digital humanities"
   ],
   "contributions": [
    {
     "title": "DaCy",
     "note": "Built an open Danish NLP pipeline on spaCy (arXiv 2107.05295).",
     "year": "2021"
    },
    {
     "title": "odyCy",
     "note": "Released an Ancient Greek spaCy pipeline.",
     "year": "2023"
    },
    {
     "title": "Danish Foundation Models paper",
     "note": "Center staff led the paper that set out the case for open Danish foundation models (arXiv 2311.07264).",
     "year": "2023"
    },
    {
     "title": "Scandinavian Embedding Benchmark",
     "note": "Co-created an open embedding benchmark for Scandinavian languages (arXiv 2406.02396).",
     "year": "2024"
    },
    {
     "title": "Danish Dynaword",
     "note": "Contributed to an openly licensed, continuously updated collection of Danish text (arXiv 2508.02271).",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "DaCy",
     "url": "https://github.com/centre-for-humanities-computing/DaCy",
     "note": "DaCy: The State of the Art Danish NLP pipeline using SpaCy  (105 stars)"
    },
    {
     "name": "tweetopic",
     "url": "https://github.com/centre-for-humanities-computing/tweetopic",
     "note": "Blazing fast topic modelling for short texts. (37 stars)"
    },
    {
     "name": "odyCy",
     "url": "https://github.com/centre-for-humanities-computing/odyCy",
     "note": "A general-purpose NLP pipeline for Ancient Greek (29 stars)"
    },
    {
     "name": "conspiracies",
     "url": "https://github.com/centre-for-humanities-computing/conspiracies",
     "note": "A python package for discovering and examining conspiracies using NLP. (10 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "chcaa/da_dacy_medium_trf",
     "url": "https://huggingface.co/chcaa/da_dacy_medium_trf",
     "kind": "model",
     "note": "Model with 6 likes on the Hub."
    },
    {
     "name": "chcaa/da_dacy_large_trf",
     "url": "https://huggingface.co/chcaa/da_dacy_large_trf",
     "kind": "model",
     "note": "Model with 5 likes on the Hub."
    },
    {
     "name": "chcaa/kb-books",
     "url": "https://huggingface.co/datasets/chcaa/kb-books",
     "kind": "dataset",
     "note": "Dataset with 5 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "DaCy: A Unified Framework for Danish NLP",
     "url": "https://arxiv.org/abs/2107.05295",
     "year": "2021"
    },
    {
     "title": "Danish Foundation Models",
     "url": "https://arxiv.org/abs/2311.07264",
     "year": "2023"
    },
    {
     "title": "The Scandinavian Embedding Benchmarks: Comprehensive Assessment of Multilingual and Monolingual Text Embedding",
     "url": "https://arxiv.org/abs/2406.02396",
     "year": "2024"
    },
    {
     "title": "Dynaword: From One-shot to Continuously Developed Datasets",
     "url": "https://arxiv.org/abs/2508.02271",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "DaCy",
     "url": "https://github.com/centre-for-humanities-computing/DaCy",
     "note": "State-of-the-art Danish NLP pipeline for spaCy"
    },
    {
     "name": "odyCy",
     "url": "https://github.com/centre-for-humanities-computing/odyCy",
     "note": "NLP pipeline for Ancient Greek"
    },
    {
     "name": "tweetopic",
     "url": "https://github.com/centre-for-humanities-computing/tweetopic",
     "note": "Topic modelling for short texts"
    },
    {
     "name": "dansk-ner",
     "url": "https://huggingface.co/datasets/chcaa/dansk-ner",
     "note": "Danish NER dataset"
    }
   ],
   "openness": "Code, models and corpora are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Kristoffer Nielbo",
     "role": "Centre director and professor",
     "url": "https://knielbo.github.io",
     "note": null
    },
    {
     "name": "Kenneth Enevoldsen",
     "role": "Assistant Professor, DaCy author",
     "url": "https://health.au.dk/en/contact/staff/show/person/kenneth.enevoldsen@cas.au.dk",
     "note": null
    },
    {
     "name": "Ross Deans Kristensen-McLachlan",
     "role": "Researcher",
     "url": "https://github.com/rdkm89",
     "note": null
    },
    {
     "name": "Márton Kardos",
     "role": "Researcher",
     "url": "https://github.com/x-tabdeveloping",
     "note": null
    },
    {
     "name": "Frida Hæstrup",
     "role": "Researcher",
     "url": "https://github.com/frillecode",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the centre website.",
   "sources": [
    "https://chc.au.dk",
    "https://huggingface.co/chcaa",
    "https://github.com/centre-for-humanities-computing",
    "https://arxiv.org/abs/2107.05295"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "center-for-humanities-computing",
   "mlat": 56.18455,
   "mlon": 10.27125,
   "logo": true
  },
  {
   "orig_id": "chai-discovery",
   "name": "Chai Discovery",
   "alias": null,
   "full_name": "Chai Discovery",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.chaidiscovery.com",
   "huggingface": "https://huggingface.co/chaidiscovery",
   "github": "https://github.com/chaidiscovery",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@chaidiscovery.com",
   "summary": "AI drug-discovery startup that open-sourced the Chai-1 biomolecular structure model.",
   "tagline": "A drug-discovery AI startup that released the open Chai-1 biomolecular structure model.",
   "about": [
    "Chai Discovery is a drug-discovery company founded in 2024 by Joshua Meier (chief executive), Jack Dent (president), Matthew McPartlon (chief technology officer) and Jacques Boitreaud. Its team came from OpenAI, Meta FAIR, Stripe and Google X, and it describes its work as predicting and reprogramming interactions between biomolecules.",
    "Chai-1, released on 9 September 2024, predicts structures of proteins, small molecules, DNA and RNA. The company published weights and inference code under Apache 2.0 and reported 77 percent success on PoseBusters against 76 percent for AlphaFold3, with competitive results even when run without multiple sequence alignments.",
    "Chai-2, announced in 2025 for de novo antibody design, is a separate model that sources describe as not openly released. Funding has included a $30 million seed round, a $70 million Series A in August 2025 led by Menlo Ventures and a $130 million Series B reported in January 2026."
   ],
   "mission": "Chai aims to let scientists design medicines computationally by predicting and engineering molecular interactions.",
   "notice": "Chai-1 is released as open source. Later models such as Chai-2 appear to be proprietary.",
   "known_for": [
    "Chai-1 open structure prediction",
    "Apache 2.0 weights and code",
    "Chai-2 antibody design"
   ],
   "focus": [
    "biology",
    "structure prediction"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in San Francisco by Meier, Dent, McPartlon and Boitreaud.",
     "year": "2024"
    },
    {
     "title": "Chai-1 released",
     "note": "Open biomolecular structure model with Apache 2.0 weights and code, plus a free web interface.",
     "year": "2024"
    },
    {
     "title": "Series A",
     "note": "A $70 million round led by Menlo Ventures, reported by Business Wire.",
     "year": "2025"
    },
    {
     "title": "Chai-2 antibody design",
     "note": "A generative model for de novo antibody design, announced in 2025.",
     "year": "2025"
    },
    {
     "title": "Series B",
     "note": "A $130 million round led by General Catalyst and Oak HC/FT, reported by GEN.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "chai-lab",
     "url": "https://github.com/chaidiscovery/chai-lab",
     "note": "Chai-1 structure prediction."
    }
   ],
   "hf_highlights": [
    {
     "name": "chai-1",
     "url": "https://huggingface.co/chaidiscovery/chai-1",
     "kind": "model",
     "note": "Chai-1 weights."
    }
   ],
   "papers": [
    {
     "title": "Chai-1: Decoding the molecular interactions of life",
     "url": "https://www.biorxiv.org/content/10.1101/2024.10.10.615955",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "chai-lab",
     "url": "https://github.com/chaidiscovery/chai-lab",
     "note": "Chai-1"
    },
    {
     "name": "chai-1",
     "url": "https://huggingface.co/chaidiscovery/chai-1",
     "note": "Weights"
    }
   ],
   "openness": "Open weights and code.",
   "key_people": [
    {
     "name": "Joshua Meier",
     "role": "Co-founder and chief executive",
     "url": "https://x.com/joshim5",
     "note": "Previously at Absci, Facebook AI and OpenAI."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.chaidiscovery.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.chaidiscovery.com",
    "https://huggingface.co/chaidiscovery"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "chai-discovery",
   "mlat": 37.7684,
   "mlon": -122.26781,
   "logo": true
  },
  {
   "orig_id": "cis-lmu",
   "name": "CIS, LMU Munich",
   "alias": null,
   "full_name": "CIS, LMU Munich",
   "type": "academic",
   "founded": 2017,
   "country": "Germany",
   "city": "Munich",
   "lat": 48.1351,
   "lon": 11.582,
   "website": "https://www.cis.uni-muenchen.de",
   "huggingface": "https://huggingface.co/cis-lmu",
   "github": "https://github.com/cisnlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Centre for Information and Language Processing at LMU Munich, behind SimAlign, Glot500 and GlotLID for massively multilingual NLP.",
   "tagline": "LMU Munich's language processing centre, behind SimAlign, Glot500 and GlotLID for massively multilingual NLP.",
   "about": [
    "The Center for Information and Language Processing at LMU Munich hosts a deep NLP group led by Hinrich Schütze, whose GitHub organization dates from 2017. Its work targets low-resource and multilingual language technology.",
    "SimAlign (2020) is a word alignment tool, Glot500 (2023) scales corpora and models to 500 languages, and GlotLID is a language identification model that covered 1665 languages in its first version and over 2,000 labels in v3.0 as of April 2024.",
    "Authors include Amir Hossein Kargaran, Ayyoob Imani and Schütze, and the data and models are published on GitHub and Hugging Face."
   ],
   "mission": "The group aims to extend language technology to many more of the world's languages and to publish the resulting data and models openly.",
   "notice": null,
   "known_for": [
    "GlotLID language identification",
    "Glot500 multilingual corpus and model",
    "SimAlign word alignment",
    "Low-resource language support"
   ],
   "focus": [
    "multilingual NLP",
    "low-resource languages",
    "language identification"
   ],
   "contributions": [
    {
     "title": "SimAlign",
     "note": "Released a word alignment tool (arXiv 2004.08728).",
     "year": "2020"
    },
    {
     "title": "Glot500",
     "note": "Scaled corpora and language models to 500 languages (arXiv 2305.12182).",
     "year": "2023"
    },
    {
     "title": "GlotLID",
     "note": "Released a language identification model for low-resource languages (arXiv 2310.16248).",
     "year": "2023"
    },
    {
     "title": "GlotLID v3.0",
     "note": "Updated the model to more than 2,000 labels.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "simalign",
     "url": "https://github.com/cisnlp/simalign",
     "note": "[EMNLP 2020] Obtain Word Alignments using Pretrained Language Models (e.g., mBERT) (397 stars)"
    },
    {
     "name": "GlotLID",
     "url": "https://github.com/cisnlp/GlotLID",
     "note": "[EMNLP 2023] 💬 Language Identification with Support for More Than 2000 Labels  (228 stars)"
    },
    {
     "name": "Glot500",
     "url": "https://github.com/cisnlp/Glot500",
     "note": "[ACL 2023] Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages (109 stars)"
    },
    {
     "name": "GlotScript",
     "url": "https://github.com/cisnlp/GlotScript",
     "note": "[LREC 2024] 🖋 Resource and Tool for Writing System Identification  (25 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "cis-lmu/glotlid",
     "url": "https://huggingface.co/cis-lmu/glotlid",
     "kind": "model",
     "note": "Model with 109 likes on the Hub."
    },
    {
     "name": "cis-lmu/glot500-base",
     "url": "https://huggingface.co/cis-lmu/glot500-base",
     "kind": "model",
     "note": "Model with 20 likes on the Hub."
    },
    {
     "name": "cis-lmu/GlotCC-V1",
     "url": "https://huggingface.co/datasets/cis-lmu/GlotCC-V1",
     "kind": "dataset",
     "note": "Dataset with 63 likes on the Hub."
    },
    {
     "name": "cis-lmu/glotlid-space",
     "url": "https://huggingface.co/spaces/cis-lmu/glotlid-space",
     "kind": "space",
     "note": "Space with 30 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages",
     "url": "https://arxiv.org/abs/2305.12182",
     "year": "2023"
    },
    {
     "title": "GlotLID: Language Identification for Low-Resource Languages",
     "url": "https://arxiv.org/abs/2310.16248",
     "year": "2023"
    },
    {
     "title": "SimAlign: High Quality Word Alignments without Parallel Training Data using Static and Contextualized Embeddings",
     "url": "https://arxiv.org/abs/2004.08728",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "GlotLID",
     "url": "https://huggingface.co/cis-lmu/glotlid",
     "note": "Language identification for over 2000 labels"
    },
    {
     "name": "Glot500",
     "url": "https://github.com/cisnlp/Glot500",
     "note": "Multilingual corpora and models for 500 languages"
    },
    {
     "name": "GlotCC",
     "url": "https://huggingface.co/datasets/cis-lmu/GlotCC-V1",
     "note": "Multilingual web corpus"
    },
    {
     "name": "SimAlign",
     "url": "https://github.com/cisnlp/simalign",
     "note": "Word alignment with pretrained language models"
    }
   ],
   "openness": "Code, models and corpora are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Hinrich Schütze",
     "role": "Professor, group head",
     "url": "https://www.cis.lmu.de/~hs/",
     "note": null
    },
    {
     "name": "Amir Hossein Kargaran",
     "role": "GlotLID lead author",
     "url": "https://kargaranamir.github.io",
     "note": null
    },
    {
     "name": "Ayyoob Imani",
     "role": "Glot500 and GlotLID author",
     "url": "https://arxiv.org/abs/2310.16248",
     "note": null
    },
    {
     "name": "Ehsaneddin Asgari",
     "role": "Former group member",
     "url": "http://llp.berkeley.edu/asgari/",
     "note": null
    },
    {
     "name": "Philipp Dufter",
     "role": "Member of the cisnlp GitHub organization",
     "url": "https://github.com/pdufter",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://www.cis.uni-muenchen.de",
    "https://huggingface.co/cis-lmu",
    "https://github.com/cisnlp",
    "https://arxiv.org/abs/2305.12182"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "cis-lmu-munich",
   "mlat": 48.0792,
   "mlon": 11.582,
   "logo": true
  },
  {
   "orig_id": "cjvt",
   "name": "CJVT",
   "alias": "University of Ljubljana",
   "full_name": "CJVT (University of Ljubljana)",
   "type": "academic",
   "founded": null,
   "country": "Slovenia",
   "city": "Ljubljana",
   "lat": 46.0569,
   "lon": 14.5058,
   "website": "https://www.cjvt.si",
   "huggingface": "https://huggingface.co/cjvt",
   "github": "https://github.com/clarinsi",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Centre for Language Resources and Technologies at the University of Ljubljana, which publishes Slovene language models, corpora and tools including the GaMS model family.",
   "tagline": "The University of Ljubljana centre that builds Slovene language resources, corpora and the open GaMS generative models.",
   "about": [
    "CJVT, the Centre for Language Resources and Technologies, is a research unit of the University of Ljubljana with its organizational base at the Faculty of Computer and Information Science and five other participating faculties. It develops and maintains digital language resources for contemporary Slovene and makes tools available free of charge.",
    "Its resources include the Sloleks morphological lexicon, which invites users to vote on accentuation and pronunciation, the ssj500k training corpus and its successor SUK, and the Tolmac system for translating lectures from Slovene. The centre joined the ELEXIS lexicography network in 2020 and works within a national program on language resources and technologies for Slovene that runs from 2019 to 2027.",
    "The GaMS (Generative Model for Slovene) family began in 2024 with a 1 billion parameter model and has grown to 2B, 9B and 27B releases built on Gemma. A March 2026 preprint presents GaMS3-12B, adapted from Gemma 3 with continual pretraining on about 140 billion tokens and an open OCR pipeline for digitizing older Slovene texts."
   ],
   "mission": "CJVT aims to build a long-term infrastructure of Slovene language resources and technology and to keep its tools and data openly available.",
   "notice": null,
   "known_for": [
    "Slovene corpora and lexicons",
    "GaMS Slovene language models",
    "Sloleks and ssj500k",
    "Open language technology for a smaller language"
   ],
   "focus": [
    "Slovene",
    "language models",
    "datasets",
    "speech"
   ],
   "contributions": [
    {
     "title": "National language technology program",
     "note": "CJVT began work in the national program on language resources and technologies for Slovene, running from 2019 to 2027.",
     "year": "2019"
    },
    {
     "title": "ELEXIS membership",
     "note": "The centre joined the European lexicography infrastructure network.",
     "year": "2020"
    },
    {
     "title": "GaMS 1B",
     "note": "Published the first Generative Model for Slovene with 1 billion parameters (arXiv 2410.06898).",
     "year": "2024"
    },
    {
     "title": "GaMS3-12B",
     "note": "Presented a 12 billion parameter Slovene model adapted from Gemma 3, with an open OCR pipeline.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "classla",
     "url": "https://github.com/clarinsi/classla",
     "note": "Fork of Stanford NLP Stanza for many languages, used in CLASSLA (53 stars)"
    },
    {
     "name": "reldi-tagger",
     "url": "https://github.com/clarinsi/reldi-tagger",
     "note": "Tagger and lemmatiser for Croatian, Serbian and Slovene (35 stars)"
    },
    {
     "name": "geobert",
     "url": "https://github.com/clarinsi/geobert",
     "note": "Geo-related BERT work (20 stars)"
    },
    {
     "name": "Slovene_ASR_e2e",
     "url": "https://github.com/clarinsi/Slovene_ASR_e2e",
     "note": "Slovene automatic speech recognition tool (20 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "cjvt/GaMS3-12B-Instruct",
     "url": "https://huggingface.co/cjvt/GaMS3-12B-Instruct",
     "kind": "model",
     "note": "12B Slovene instruct model, the most downloaded CJVT model."
    },
    {
     "name": "cjvt/GaMS-9B-Instruct",
     "url": "https://huggingface.co/cjvt/GaMS-9B-Instruct",
     "kind": "model",
     "note": "9B Slovene instruct model."
    },
    {
     "name": "cjvt/slovenian-llm-eval",
     "url": "https://huggingface.co/datasets/cjvt/slovenian-llm-eval",
     "kind": "dataset",
     "note": "Evaluation data for Slovenian LLMs."
    },
    {
     "name": "cjvt/sloleks",
     "url": "https://huggingface.co/datasets/cjvt/sloleks",
     "kind": "dataset",
     "note": "Slovene lexicon data."
    }
   ],
   "papers": [
    {
     "title": "Generative Model for Less-Resourced Language with 1 billion parameters",
     "url": "https://arxiv.org/abs/2410.06898",
     "year": "2024"
    },
    {
     "title": "Building a Strong Instruction Language Model for a Less-Resourced Language",
     "url": "https://arxiv.org/abs/2603.01691",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "GaMS3-12B-Instruct",
     "url": "https://huggingface.co/cjvt/GaMS3-12B-Instruct",
     "note": "Slovene instruction model"
    },
    {
     "name": "t5-sl-large",
     "url": "https://huggingface.co/cjvt/t5-sl-large",
     "note": "Slovene T5 model"
    },
    {
     "name": "Slovene_ASR_e2e",
     "url": "https://github.com/clarinsi/Slovene_ASR_e2e",
     "note": "Slovene ASR tool"
    }
   ],
   "openness": "Models and corpora are public on Hugging Face and GitHub, with licences set per project.",
   "key_people": [
    {
     "name": "Simon Krek",
     "role": "Centre contact and corpus author",
     "url": "https://www.cjvt.si/en/",
     "note": "Listed as contact on the CJVT site and author of Slovene corpora."
    },
    {
     "name": "Marko Robnik-Sikonja",
     "role": "Professor, GaMS lead author",
     "url": "https://arxiv.org/abs/2603.01691",
     "note": "Co-author of the GaMS papers at the Faculty of Computer and Information Science."
    },
    {
     "name": "Domen Vres",
     "role": "Researcher, GaMS first author",
     "url": "https://arxiv.org/abs/2410.06898",
     "note": "First author of the GaMS 1B paper."
    }
   ],
   "discord_note": null,
   "contact_note": "The CJVT website lists organisation, staff and news pages; its GitHub presence is the clarinsi organisation.",
   "sources": [
    "https://www.cjvt.si",
    "https://huggingface.co/cjvt",
    "https://github.com/clarinsi"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "cjvt",
   "mlat": 46.02154,
   "mlon": 14.5058,
   "logo": true
  },
  {
   "orig_id": "ckip-lab",
   "name": "CKIP Lab",
   "alias": "Academia Sinica",
   "full_name": "CKIP Lab (Academia Sinica)",
   "type": "academic",
   "founded": 1986,
   "country": "Taiwan",
   "city": "Taipei",
   "lat": 25.03,
   "lon": 121.57,
   "website": "https://ckip.iis.sinica.edu.tw/",
   "huggingface": "https://huggingface.co/ckiplab",
   "github": "https://github.com/ckiplab",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Non-profit NLP lab publishing Chinese NLP models including Han Chinese models.",
   "tagline": "An Academia Sinica research group, active since 1986, that builds Chinese language processing resources and publishes open Chinese NLP models.",
   "about": [
    "CKIP, the Chinese Knowledge and Information Processing group, is a cross-institute research team of Academia Sinica's Institute of Information Science and Institute of Linguistics. The group dates itself to 1986 and aims to supply resources and research environments for Chinese natural language processing.",
    "Its outputs include word segmentation, parsing, a sentence structure treebank, the Sinica Corpus, the E-HowNet lexical knowledge base, a Kangxi Dictionary ontology, language models and Chinese embeddings. Current research covers deep learning, knowledge representation, language understanding and knowledge extraction.",
    "CKIP Transformers provides traditional Chinese ALBERT, BERT and GPT-2 models with tools for word segmentation, part-of-speech tagging and named entity recognition, distributed through GitHub, Hugging Face and PyPI under GPL-3.0. Wei-Yun Ma, an associate research fellow, is principal investigator of the lab."
   ],
   "mission": "CKIP works to provide computational resources and research environments for Chinese language processing.",
   "notice": null,
   "known_for": [
    "Chinese word segmentation and parsing tools",
    "Sinica Corpus and E-HowNet",
    "Open Chinese language models"
   ],
   "focus": [
    "Chinese NLP",
    "language models"
   ],
   "contributions": [
    {
     "title": "Group founded",
     "note": "CKIP was established at Academia Sinica.",
     "year": "1986"
    }
   ],
   "github_highlights": [
    {
     "name": "ckiptagger",
     "url": "https://github.com/ckiplab/ckiptagger",
     "note": "Chinese word segmentation, POS and NER."
    },
    {
     "name": "ckip-transformers",
     "url": "https://github.com/ckiplab/ckip-transformers",
     "note": "Transformers models."
    },
    {
     "name": "ckipnlp",
     "url": "https://github.com/ckiplab/ckipnlp",
     "note": "CoreNLP toolkit."
    },
    {
     "name": "ehownet",
     "url": "https://github.com/ckiplab/ehownet",
     "note": "E-HowNet utilities."
    }
   ],
   "hf_highlights": [
    {
     "name": "ckiplab/bert-base-chinese-ner",
     "url": "https://huggingface.co/ckiplab/bert-base-chinese-ner",
     "kind": "model",
     "note": "NER."
    },
    {
     "name": "ckiplab/bert-base-chinese-ws",
     "url": "https://huggingface.co/ckiplab/bert-base-chinese-ws",
     "kind": "model",
     "note": "Word segmentation."
    },
    {
     "name": "ckiplab/bert-base-chinese-pos",
     "url": "https://huggingface.co/ckiplab/bert-base-chinese-pos",
     "kind": "model",
     "note": "POS."
    },
    {
     "name": "ckiplab/albert-tiny-chinese-ws",
     "url": "https://huggingface.co/ckiplab/albert-tiny-chinese-ws",
     "kind": "model",
     "note": "Small WS."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "bert-base-han-chinese-pos-xiandai",
     "url": "https://huggingface.co/ckiplab/bert-base-han-chinese-pos-xiandai",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Models on Hugging Face and code on GitHub.",
   "key_people": [
    {
     "name": "Wei-Yun Ma",
     "role": "Principal investigator, CKIP NLP lab",
     "url": "https://homepage.iis.sinica.edu.tw/pages/ma/vita_en.html",
     "note": "Associate research fellow at the Institute of Information Science; PhD from Columbia University."
    }
   ],
   "discord_note": null,
   "contact_note": "ckip.iis.sinica.edu.tw is the lab homepage.",
   "sources": [
    "https://huggingface.co/ckiplab",
    "https://ckip.iis.sinica.edu.tw/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "ckip-lab",
   "mlat": 24.99464,
   "mlon": 121.57,
   "logo": true
  },
  {
   "orig_id": "clarin-pl",
   "name": "CLARIN-PL",
   "alias": "Wroclaw",
   "full_name": "CLARIN-PL (Wroclaw)",
   "type": "academic",
   "founded": 2015,
   "country": "Poland",
   "city": "Wroclaw",
   "lat": 51.1079,
   "lon": 17.0385,
   "website": "https://clarin-pl.eu",
   "huggingface": "https://huggingface.co/clarin-pl",
   "github": "https://github.com/CLARIN-PL",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "CLARIN-PL is the Polish language technology infrastructure based at Wroclaw University of Science and Technology, publishing Polish NLP tools, models, benchmarks and corpora.",
   "tagline": "Polish language technology infrastructure at Wrocław University of Science and Technology, publishing Polish NLP tools, corpora and benchmarks.",
   "about": [
    "CLARIN-PL is the Polish node of the CLARIN research infrastructure, based at Wrocław University of Science and Technology. Its GitHub organization dates from 2015.",
    "Its repository hosts plWordNet, led by Maciej Piasecki, with versions 3.0 and 4.0 reaching hundreds of thousands of senses and a licence allowing free use. It also provides the KPWr corpus, the WordnetLoom editor, embeddings tools and the LEPISZCZE benchmark.",
    "The resources serve linguists, humanities scholars and NLP developers working on Polish."
   ],
   "mission": "CLARIN-PL provides language resources and tools for Polish humanities and social science research and shares them through an open repository.",
   "notice": null,
   "known_for": [
    "plWordNet Polish wordnet",
    "LEPISZCZE Polish benchmark",
    "KPWr corpus",
    "CLARIN repository for Polish"
   ],
   "focus": [
    "Polish NLP",
    "benchmarks"
   ],
   "contributions": [
    {
     "title": "plWordNet",
     "note": "The Polish wordnet was published in the CLARIN-PL repository with Piasecki as lead author.",
     "year": "2010"
    },
    {
     "title": "embeddings library",
     "note": "Released a Polish embeddings toolkit repository.",
     "year": "2021"
    },
    {
     "title": "plWordNet to Princeton mapping",
     "note": "Rudnicka, Witkowski and Piasecki described a large-scale mapping to Princeton WordNet.",
     "year": "2021"
    },
    {
     "title": "LEPISZCZE",
     "note": "Designed a Polish NLP benchmark (arXiv 2211.13112).",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "embeddings",
     "url": "https://github.com/CLARIN-PL/embeddings",
     "note": "Embeddings: State-of-the-art Text Representations for Natural Language Processing tasks, an initial version of (37 stars)"
    },
    {
     "name": "PolDeepNer",
     "url": "https://github.com/CLARIN-PL/PolDeepNer",
     "note": "Tool for named entity recognition for Polish based on deep learning. (31 stars)"
    },
    {
     "name": "chatgpt-evaluation-01-2023",
     "url": "https://github.com/CLARIN-PL/chatgpt-evaluation-01-2023",
     "note": "Code, datasets and results of the ChatGPT evaluation presented in paper \"ChatGPT: Jack of all trades, master o (28 stars)"
    },
    {
     "name": "LEPISZCZE",
     "url": "https://github.com/CLARIN-PL/LEPISZCZE",
     "note": "This is the way: designing and compiling LEPISZCZE, a comprehensive NLP benchmark for Polish (15 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "clarin-pl/FastPDN",
     "url": "https://huggingface.co/clarin-pl/FastPDN",
     "kind": "model",
     "note": "Model with 12 likes on the Hub."
    },
    {
     "name": "clarin-pl/fastText-kgr10",
     "url": "https://huggingface.co/clarin-pl/fastText-kgr10",
     "kind": "model",
     "note": "Model with 2 likes on the Hub."
    },
    {
     "name": "clarin-pl/kpwr-ner",
     "url": "https://huggingface.co/datasets/clarin-pl/kpwr-ner",
     "kind": "dataset",
     "note": "Dataset with 10 likes on the Hub."
    },
    {
     "name": "clarin-pl/datasets-explorer",
     "url": "https://huggingface.co/spaces/clarin-pl/datasets-explorer",
     "kind": "space",
     "note": "Space with 2 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "This is the way: designing and compiling LEPISZCZE, a comprehensive NLP benchmark for Polish",
     "url": "https://arxiv.org/abs/2211.13112",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "LEPISZCZE",
     "url": "https://github.com/CLARIN-PL/LEPISZCZE",
     "note": "Polish NLP benchmark"
    },
    {
     "name": "embeddings",
     "url": "https://github.com/CLARIN-PL/embeddings",
     "note": "Text representation library for NLP tasks"
    },
    {
     "name": "PolDeepNer",
     "url": "https://github.com/CLARIN-PL/PolDeepNer",
     "note": "Deep learning NER for Polish"
    },
    {
     "name": "FastPDN",
     "url": "https://huggingface.co/clarin-pl/FastPDN",
     "note": "Named entity recogniser model"
    }
   ],
   "openness": "Code, models and datasets are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Maciej Piasecki",
     "role": "Lead author of plWordNet",
     "url": "https://clarin-pl.eu/dspace/handle/11321/29",
     "note": null
    },
    {
     "name": "Danijel Koržinek",
     "role": "Member of the CLARIN-PL GitHub organization",
     "url": "https://github.com/danijel3",
     "note": null
    },
    {
     "name": "Jan Kocoń",
     "role": "Member of the CLARIN-PL GitHub organization",
     "url": "https://github.com/KoconJan",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://clarin-pl.eu",
    "https://huggingface.co/clarin-pl",
    "https://github.com/CLARIN-PL",
    "https://arxiv.org/abs/2211.13112"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "clarin-pl",
   "logo": true
  },
  {
   "orig_id": "classla",
   "name": "CLASSLA",
   "alias": "CLARIN Knowledge Centre for South Slavic languages",
   "full_name": "CLASSLA (CLARIN Knowledge Centre for South Slavic languages)",
   "type": "institute",
   "founded": 2019,
   "country": "Slovenia",
   "city": "Ljubljana",
   "lat": 46.0569,
   "lon": 14.5058,
   "website": "https://www.clarin.si/info/k-centre/",
   "huggingface": "https://huggingface.co/classla",
   "github": "https://github.com/clarinsi",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "helpdesk.classla@clarin.si",
   "summary": "CLARIN knowledge centre providing open tools, models and corpora for Slovenian, Croatian, Serbian, Bosnian and Montenegrin.",
   "tagline": "CLARIN knowledge centre providing open tools, models and corpora for Slovenian, Croatian, Serbian, Bosnian and Montenegrin.",
   "about": [
    "CLASSLA was recognised by CLARIN on 19 March 2019. It is operated by CLARIN.SI, the Institute of Croatian Language and CLADA-BG, with hosting at the Jožef Stefan Institute in Ljubljana.",
    "It offers a helpdesk, the CLASSLA-Stanza pipeline for South Slavic languages, with an improved version released in September 2023, and the largest general web corpora for these languages. BERTić, a model for BCMS languages, is described in arXiv 2104.09243.",
    "Recent work includes benchmarks and an LLM evaluation dashboard for South Slavic languages."
   ],
   "mission": "The centre provides expertise and open resources for processing South Slavic languages and supports researchers across the region.",
   "notice": null,
   "known_for": [
    "CLASSLA-Stanza pipeline",
    "South Slavic web corpora",
    "BERTić BCMS model",
    "South Slavic LLM benchmarks"
   ],
   "focus": [
    "South Slavic NLP",
    "language models",
    "corpora"
   ],
   "contributions": [
    {
     "title": "Recognised as knowledge centre",
     "note": "CLARIN officially recognised CLASSLA.",
     "year": "2019"
    },
    {
     "title": "CLASSLA-Stanza",
     "note": "The classla repository for linguistic processing was created.",
     "year": "2019"
    },
    {
     "title": "BERTić",
     "note": "Released a language model for BCMS languages (arXiv 2104.09243).",
     "year": "2021"
    },
    {
     "title": "Improved pipeline",
     "note": "Released a new state-of-the-art CLASSLA-Stanza version.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "reldi-tagger",
     "url": "https://github.com/clarinsi/reldi-tagger",
     "note": "A tagger and lemmatiser for Croatian, Serbian and Slovene. (35 stars)"
    },
    {
     "name": "geobert",
     "url": "https://github.com/clarinsi/geobert",
     "note": "Open-source repository. (20 stars)"
    },
    {
     "name": "Slovene_ASR_e2e",
     "url": "https://github.com/clarinsi/Slovene_ASR_e2e",
     "note": "Automatic Speech Recognition tool (20 stars)"
    },
    {
     "name": "csmtiser",
     "url": "https://github.com/clarinsi/csmtiser",
     "note": "A tool for text normalisation via character-level machine translation (13 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "classla/xlm-roberta-base-multilingual-text-genre-classifier",
     "url": "https://huggingface.co/classla/xlm-roberta-base-multilingual-text-genre-classifier",
     "kind": "model",
     "note": "Model with 31 likes on the Hub."
    },
    {
     "name": "classla/multilingual-IPTC-news-topic-classifier",
     "url": "https://huggingface.co/classla/multilingual-IPTC-news-topic-classifier",
     "kind": "model",
     "note": "Model with 22 likes on the Hub."
    },
    {
     "name": "classla/ParlaSent",
     "url": "https://huggingface.co/datasets/classla/ParlaSent",
     "kind": "dataset",
     "note": "Dataset with 7 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "BERTić -- The Transformer Language Model for Bosnian, Croatian, Montenegrin and Serbian",
     "url": "https://arxiv.org/abs/2104.09243",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "BERTic",
     "url": "https://huggingface.co/classla/bcms-bertic",
     "note": "Transformer for Bosnian, Croatian, Montenegrin and Serbian"
    },
    {
     "name": "IPTC news topic classifier",
     "url": "https://huggingface.co/classla/multilingual-IPTC-news-topic-classifier",
     "note": "Multilingual news classifier"
    },
    {
     "name": "ParlaSpeech",
     "url": "https://huggingface.co/datasets/classla/ParlaSpeech-HR",
     "note": "Parliamentary speech dataset"
    },
    {
     "name": "classla",
     "url": "https://github.com/clarinsi/classla",
     "note": "Stanza fork for South Slavic languages"
    }
   ],
   "openness": "Tools, models and datasets are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Nikola Ljubešić",
     "role": "Author of CLASSLA papers",
     "url": "https://www.clarin.eu/blog/tour-de-clarin-clarin-knowledge-centre-south-slavic-languages-classla",
     "note": null
    },
    {
     "name": "Taja Kuzman Pungeršek",
     "role": "Researcher",
     "url": "https://www.clarin.eu/blog/tour-de-clarin-clarin-knowledge-centre-south-slavic-languages-classla",
     "note": null
    },
    {
     "name": "Tomaž Erjavec",
     "role": "Author of CLASSLA overview",
     "url": "https://www.clarin.eu/blog/tour-de-clarin-clarin-knowledge-centre-south-slavic-languages-classla",
     "note": null
    },
    {
     "name": "Petya Osenova",
     "role": "Author, CLADA-BG",
     "url": "https://www.clarin.eu/blog/tour-de-clarin-clarin-knowledge-centre-south-slavic-languages-classla",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the helpdesk address printed on the site.",
   "sources": [
    "https://www.clarin.si/info/k-centre/",
    "https://huggingface.co/classla",
    "https://github.com/clarinsi",
    "https://arxiv.org/abs/2104.09243"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "classla",
   "mlat": 46.09226,
   "mlon": 14.5058,
   "logo": true
  },
  {
   "orig_id": "clips-antwerp",
   "name": "CLiPS",
   "alias": "University of Antwerp",
   "full_name": "CLiPS (University of Antwerp)",
   "type": "academic",
   "founded": 2011,
   "country": "Belgium",
   "city": "Antwerp",
   "lat": 51.2194,
   "lon": 4.4025,
   "website": "https://www.uantwerpen.be/en/research-groups/clips/",
   "huggingface": "https://huggingface.co/clips",
   "github": "https://github.com/clips",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computational linguistics group in Antwerp, creators of the Pattern web-mining library and Dutch embedding and FAQ resources.",
   "tagline": "University of Antwerp computational linguistics group, creator of the Pattern web mining library and Dutch NLP resources.",
   "about": [
    "CLiPS is the computational linguistics and psycholinguistics research centre at the University of Antwerp, with a GitHub organization created in 2011.",
    "Pattern, by Tom De Smedt and Walter Daelemans, is a BSD-licensed Python package for web mining and NLP with Dutch modules, described in a JMLR 2012 paper. The repository holds about 8,800 stars.",
    "The group also released the multilingual FAQ dataset MFAQ and Dutch embedding models."
   ],
   "mission": "CLiPS develops computational models of language and shares software and datasets, with attention to Dutch.",
   "notice": null,
   "known_for": [
    "Pattern web mining library",
    "Dutch language tools",
    "MFAQ dataset",
    "Dutch embeddings"
   ],
   "focus": [
    "Dutch NLP",
    "embeddings",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Pattern repository",
     "note": "The Pattern web mining module was published on GitHub.",
     "year": "2011"
    },
    {
     "title": "Pattern JMLR paper",
     "note": "De Smedt and Daelemans described Pattern, a BSD-licensed Python package with Dutch modules, in JMLR.",
     "year": "2012"
    },
    {
     "title": "Creative web services with Pattern",
     "note": "De Smedt, Nijs and Daelemans published a follow-up on applications built on Pattern.",
     "year": "2014"
    },
    {
     "title": "MFAQ",
     "note": "Released a multilingual FAQ dataset (arXiv 2109.12870).",
     "year": "2021"
    }
   ],
   "github_highlights": [
    {
     "name": "pattern",
     "url": "https://github.com/clips/pattern",
     "note": "Web mining module for Python, with tools for scraping, natural language processing, machine learning, network  (8,857 stars)"
    },
    {
     "name": "clicr",
     "url": "https://github.com/clips/clicr",
     "note": "Machine reading comprehension on clinical case reports (154 stars)"
    },
    {
     "name": "news-audit",
     "url": "https://github.com/clips/news-audit",
     "note": "Fake news detection, Google Summer of Code 2017 (92 stars)"
    },
    {
     "name": "dutchembeddings",
     "url": "https://github.com/clips/dutchembeddings",
     "note": "Repository for the word embeddings experiments described in \"Evaluating Unsupervised Dutch Word Embeddings as  (86 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "clips/mfaq",
     "url": "https://huggingface.co/clips/mfaq",
     "kind": "model",
     "note": "Model with 35 likes on the Hub."
    },
    {
     "name": "clips/e5-large-trm-nl",
     "url": "https://huggingface.co/clips/e5-large-trm-nl",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "clips/mqa",
     "url": "https://huggingface.co/datasets/clips/mqa",
     "kind": "dataset",
     "note": "Dataset with 58 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "MFAQ: a Multilingual FAQ Dataset",
     "url": "https://arxiv.org/abs/2109.12870",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Pattern",
     "url": "https://github.com/clips/pattern",
     "note": "Web mining module for Python"
    },
    {
     "name": "MQA",
     "url": "https://huggingface.co/datasets/clips/mqa",
     "note": "Multilingual QA dataset"
    },
    {
     "name": "MFAQ",
     "url": "https://huggingface.co/datasets/clips/mfaq",
     "note": "Multilingual FAQ dataset"
    },
    {
     "name": "e5 Dutch models",
     "url": "https://huggingface.co/clips/e5-base-trm-nl",
     "note": "Dutch E5 embeddings"
    }
   ],
   "openness": "Code, datasets and models are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Tom De Smedt",
     "role": "Author of Pattern",
     "url": "https://jmlr2020.csail.mit.edu/papers/volume13/desmedt12a/desmedt12a.pdf",
     "note": "First author of the 2012 JMLR paper on Pattern."
    },
    {
     "name": "Walter Daelemans",
     "role": "Professor and research director of CLiPS",
     "url": "https://www.clips.uantwerpen.be/~walter/",
     "note": "Full professor of computational linguistics at Antwerp since 1999; founded the ILK group at Tilburg."
    },
    {
     "name": "Enrique Manjavacas",
     "role": "Member of the clips GitHub organization",
     "url": "https://emanjavacas.github.io",
     "note": null
    },
    {
     "name": "Stephan Tulkens",
     "role": "Member of the clips GitHub organization",
     "url": "http://stephantul.github.io/",
     "note": null
    },
    {
     "name": "Pieter Fivez",
     "role": "Member of the clips GitHub organization",
     "url": "https://www.uantwerpen.be/en/core-facilities/textua/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://www.uantwerpen.be/en/research-groups/clips/",
    "https://huggingface.co/clips",
    "https://github.com/clips",
    "https://arxiv.org/abs/2109.12870"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "clips",
   "logo": true
  },
  {
   "orig_id": "cltl-vu",
   "name": "CLTL",
   "alias": "VU Amsterdam",
   "full_name": "CLTL (VU Amsterdam)",
   "type": "academic",
   "founded": null,
   "country": "Netherlands",
   "city": "Amsterdam",
   "lat": 52.3676,
   "lon": 4.9041,
   "website": "https://cltl.nl",
   "huggingface": "https://huggingface.co/CLTL",
   "github": "https://github.com/cltl",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computational Linguistics and Text Mining Lab at Vrije Universiteit Amsterdam, with open teaching material, Dutch wordnet resources and a Dutch medical language model.",
   "tagline": "VU Amsterdam's Computational Linguistics and Text Mining Lab, publishing open Dutch language resources, MedRoBERTa.nl and teaching material.",
   "about": [
    "The Computational Linguistics and Text Mining Lab at Vrije Universiteit Amsterdam is led by Piek Vossen and Antske Fokkens. It has moved from computational lexicology and terminology toward broader work in AI and computational linguistics, with research themes in identity, reference and perspectives, meaning representation, and large language models and interpretability.",
    "Vossen founded the Global WordNet Association in 2000 and presided over it until 2024, and the lab contributes to the Dutch Global Wordnet Grid. It states that it publishes models, software and resources as open source, and also teaches a Language and AI master's program.",
    "MedRoBERTa.nl, introduced in 2021 by Stella Verkijk and Vossen, was pre-trained on nearly 10 million hospital notes from Amsterdam UMC. The lab's Python for Text Analysis course, created in 2016, has over 450 GitHub stars."
   ],
   "mission": "The lab combines computational linguistics and text mining and releases its models, software and teaching resources openly.",
   "notice": null,
   "known_for": [
    "MedRoBERTa.nl clinical Dutch model",
    "Dutch wordnet work",
    "Python for text analysis course",
    "Open lab software"
   ],
   "focus": [
    "Dutch NLP"
   ],
   "contributions": [
    {
     "title": "Global WordNet Association",
     "note": "Vossen founded the association that coordinates wordnets across languages.",
     "year": "2000"
    },
    {
     "title": "Python for Text Analysis",
     "note": "The open teaching course repository was created.",
     "year": "2016"
    },
    {
     "title": "MedRoBERTa.nl",
     "note": "Released a Dutch clinical language model trained on hospital notes.",
     "year": "2021"
    },
    {
     "title": "Medical model paper",
     "note": "An Artificial Intelligence in Medicine article described the model's development and anonymization.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "python-for-text-analysis",
     "url": "https://github.com/cltl/python-for-text-analysis",
     "note": "If you want to use Python for text analysis, this course is for you! (457 stars)"
    },
    {
     "name": "OpenDutchWordnet",
     "url": "https://github.com/cltl/OpenDutchWordnet",
     "note": "This repo provides a python module to work with Open Dutch WordNet.  It was created using python 3.4. (70 stars)"
    },
    {
     "name": "pepper",
     "url": "https://github.com/cltl/pepper",
     "note": "VU-CLTL Pepper/Nao Application Repository (Python 2) (31 stars)"
    },
    {
     "name": "wsd-dynamic-sense-vector",
     "url": "https://github.com/cltl/wsd-dynamic-sense-vector",
     "note": "Open-source repository. (25 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "CLTL/MedRoBERTa.nl",
     "url": "https://huggingface.co/CLTL/MedRoBERTa.nl",
     "kind": "model",
     "note": "Model with 15 likes on the Hub."
    },
    {
     "name": "CLTL/icf-domains",
     "url": "https://huggingface.co/CLTL/icf-domains",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "CLTL/GoEmotions_mBERT",
     "url": "https://huggingface.co/spaces/CLTL/GoEmotions_mBERT",
     "kind": "space",
     "note": "Space with 0 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "python-for-text-analysis",
     "url": "https://github.com/cltl/python-for-text-analysis",
     "note": "Popular open course on Python for text analysis"
    },
    {
     "name": "OpenDutchWordnet",
     "url": "https://github.com/cltl/OpenDutchWordnet",
     "note": "Python module for the Open Dutch WordNet"
    },
    {
     "name": "MedRoBERTa.nl",
     "url": "https://huggingface.co/CLTL/MedRoBERTa.nl",
     "note": "Dutch medical RoBERTa model"
    },
    {
     "name": "EventStoryLine",
     "url": "https://github.com/cltl/EventStoryLine",
     "note": "Corpus for storyline extraction"
    }
   ],
   "openness": "Courses, tools, datasets and models are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Piek Vossen",
     "role": "Professor of Computational Lexicology; lab founder and head",
     "url": "https://vossen.info",
     "note": "Spinoza laureate (2013); founder of the Global WordNet Association."
    },
    {
     "name": "Antske Fokkens",
     "role": "Professor and co-lead of the lab",
     "url": "https://vu.nl/en/research/scientists/antske-fokkens",
     "note": null
    },
    {
     "name": "Stella Verkijk",
     "role": "MedRoBERTa.nl author",
     "url": "https://research.vu.nl/en/publications/medrobertanl-a-language-model-for-dutch-electronic-health-records/",
     "note": null
    },
    {
     "name": "Lea Krause",
     "role": "Member of the cltl GitHub organization",
     "url": "https://lkra.github.io",
     "note": null
    },
    {
     "name": "Stefan Schouten",
     "role": "Member of the cltl GitHub organization",
     "url": "https://sfschouten.github.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://cltl.nl",
    "https://huggingface.co/CLTL",
    "https://github.com/cltl"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "cltl",
   "logo": true
  },
  {
   "orig_id": "common-crawl",
   "name": "Common Crawl Foundation",
   "alias": null,
   "full_name": "Common Crawl Foundation",
   "type": "nonprofit",
   "founded": 2007,
   "country": "United States",
   "city": "Beverly Hills",
   "lat": 34.0736,
   "lon": -118.4004,
   "website": "https://commoncrawl.org",
   "huggingface": "https://huggingface.co/commoncrawl",
   "github": "https://github.com/commoncrawl",
   "x": null,
   "discord": "https://discord.gg/njaVFh7avF",
   "linkedin": null,
   "contact_email": null,
   "summary": "Nonprofit that publishes a free, open repository of web crawl data used to train many language models.",
   "tagline": "A nonprofit that has published free, open web crawl archives since 2008, forming a foundation dataset for language model research.",
   "about": [
    "Common Crawl is a 501(c)(3) nonprofit founded in 2007 by Gil Elbaz, who had earlier co-founded Applied Semantics. Its premise was that web-scale crawl data should be available to researchers who lack the resources to run their own crawling infrastructure.",
    "The foundation publishes monthly crawl archives, each typically containing more than two billion pages, in WARC, WAT and WET formats. It also releases URL and CDXJ indexes and host-level and domain-level web graphs. The data is hosted on Amazon Web Services and can be downloaded free of charge.",
    "A small staff of engineers and researchers maintains the crawler and the datasets, supported by a board and an advisory group. The archives serve academic researchers, startups and the teams that build language models."
   ],
   "mission": "Common Crawl aims to keep a large, current copy of the public web freely available so that anyone can analyze it. It treats open access to crawl data as a way to widen who can do research on the web and on language technology.",
   "notice": null,
   "known_for": [
    "Open monthly web crawl archives",
    "More than 10 petabytes of crawled data since 2008",
    "Web graphs with PageRank and harmonic centrality",
    "A common source of training text for language models"
   ],
   "focus": [
    "datasets",
    "web crawling"
   ],
   "contributions": [
    {
     "title": "Foundation established",
     "note": "Gil Elbaz founded the nonprofit to publish crawl data for public use.",
     "year": "2007"
    },
    {
     "title": "First archive crawl",
     "note": "Crawl archives were made available publicly beginning in 2008.",
     "year": "2008"
    },
    {
     "title": "Web graph releases",
     "note": "Host-level and domain-level web graphs with Harmonic Centrality and PageRank values are published alongside the crawls.",
     "year": "2020"
    }
   ],
   "github_highlights": [
    {
     "name": "commoncrawl",
     "url": "https://github.com/commoncrawl/commoncrawl",
     "note": "Crawl support library."
    },
    {
     "name": "cc-pyspark",
     "url": "https://github.com/commoncrawl/cc-pyspark",
     "note": "PySpark."
    },
    {
     "name": "news-crawl",
     "url": "https://github.com/commoncrawl/news-crawl",
     "note": "News crawl."
    },
    {
     "name": "cc-crawl-statistics",
     "url": "https://github.com/commoncrawl/cc-crawl-statistics",
     "note": "Statistics."
    }
   ],
   "hf_highlights": [
    {
     "name": "CommonLID",
     "url": "https://huggingface.co/datasets/commoncrawl/CommonLID",
     "kind": "dataset",
     "note": "Language ID."
    },
    {
     "name": "statistics",
     "url": "https://huggingface.co/datasets/commoncrawl/statistics",
     "kind": "dataset",
     "note": "Statistics."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Crawl archive",
     "url": "https://commoncrawl.org",
     "note": "Open crawl data since 2008"
    },
    {
     "name": "cc-pyspark",
     "url": "https://github.com/commoncrawl/cc-pyspark",
     "note": "Processing examples"
    }
   ],
   "openness": "Data free to use; code open source.",
   "key_people": [
    {
     "name": "Rich Skrenta",
     "role": "Executive Director",
     "url": "https://commoncrawl.org/team",
     "note": null
    },
    {
     "name": "Gil Elbaz",
     "role": "Founder and Chairman",
     "url": "https://commoncrawl.org/team",
     "note": "Founded Common Crawl in 2007."
    },
    {
     "name": "Greg Lindahl",
     "role": "Chief Technology Officer",
     "url": "https://commoncrawl.org/team",
     "note": null
    },
    {
     "name": "Sebastian Nagel",
     "role": "Distinguished Engineer",
     "url": "https://commoncrawl.org/team",
     "note": null
    },
    {
     "name": "Thom Vaughan",
     "role": "Principal Engineer",
     "url": "https://commoncrawl.org/team",
     "note": null
    },
    {
     "name": "Pedro Ortiz Suarez",
     "role": "Principal Research Scientist",
     "url": "https://commoncrawl.org/team",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://commoncrawl.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://commoncrawl.org",
    "https://commoncrawl.org/contact-us",
    "https://github.com/commoncrawl"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "common-crawl-foundation",
   "logo": true
  },
  {
   "orig_id": "compvis",
   "name": "CompVis",
   "alias": "LMU Munich",
   "full_name": "CompVis (LMU Munich)",
   "type": "academic",
   "founded": null,
   "country": "Germany",
   "city": "Munich",
   "lat": 48.1351,
   "lon": 11.582,
   "website": "https://ommer-lab.com",
   "huggingface": "https://huggingface.co/CompVis",
   "github": "https://github.com/CompVis",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computer Vision and Learning group at LMU Munich, creators of latent diffusion and the research behind Stable Diffusion. It continues to publish open code and models.",
   "tagline": "The Computer Vision and Learning group at LMU Munich, which developed latent diffusion, the research behind Stable Diffusion, and releases its code openly.",
   "about": [
    "CompVis is the Computer Vision and Learning group at LMU Munich, formerly the Computer Vision Group at Heidelberg University, led by Prof. Dr. Bjorn Ommer. Its research covers generative visual synthesis, invertible deep models for explainable AI, deep metric and representation learning, and self-supervised learning.",
    "The group is best known for latent diffusion models, presented at CVPR 2022 as a method for high-resolution image synthesis. Source code and models were released on GitHub, and the approach underlies Stable Diffusion.",
    "Newer projects listed on the group's site include SCFlow, TREAD and CleanDIFT. The group also applies its methods to the digital humanities and the life sciences."
   ],
   "mission": "The group pursues fundamental research in computer vision and machine learning, with an emphasis on generative models and interpretable representations. It publishes code and models so that others can build on the work.",
   "notice": null,
   "known_for": [
    "Latent diffusion models",
    "The research basis of Stable Diffusion",
    "Open image synthesis code and models",
    "Explainable and generative vision research"
   ],
   "focus": [
    "vision",
    "generative models",
    "diffusion"
   ],
   "contributions": [
    {
     "title": "Taming Transformers",
     "note": "Combined convolutional and transformer models for high-resolution synthesis, with code released.",
     "year": "2021"
    },
    {
     "title": "Latent diffusion models",
     "note": "The group presented high-resolution image synthesis with latent diffusion at CVPR and released code and models on GitHub.",
     "year": "2022"
    },
    {
     "title": "Stable Diffusion",
     "note": "The group's latent diffusion work was the foundation for the Stable Diffusion image generator.",
     "year": "2022"
    },
    {
     "title": "CleanDIFT",
     "note": "Released a method for extracting diffusion features without added noise.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "stable-diffusion",
     "url": "https://github.com/CompVis/stable-diffusion",
     "note": "Latent text-to-image diffusion model (73,576 stars)"
    },
    {
     "name": "latent-diffusion",
     "url": "https://github.com/CompVis/latent-diffusion",
     "note": "High-resolution image synthesis with latent diffusion models (14,190 stars)"
    },
    {
     "name": "taming-transformers",
     "url": "https://github.com/CompVis/taming-transformers",
     "note": "VQGAN and transformers for image synthesis (6,533 stars)"
    },
    {
     "name": "depth-fm",
     "url": "https://github.com/CompVis/depth-fm",
     "note": "Fast monocular depth estimation with flow matching (764 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "CompVis/stable-diffusion-v1-4",
     "url": "https://huggingface.co/CompVis/stable-diffusion-v1-4",
     "kind": "model",
     "note": "Most downloaded model of the group, the Stable Diffusion v1.4 weights."
    },
    {
     "name": "CompVis/stable-diffusion-safety-checker",
     "url": "https://huggingface.co/CompVis/stable-diffusion-safety-checker",
     "kind": "model",
     "note": "Safety checker component used by Stable Diffusion pipelines."
    },
    {
     "name": "CompVis/cleandift",
     "url": "https://huggingface.co/CompVis/cleandift",
     "kind": "model",
     "note": "Diffusion features without noise, from the CleanDIFT paper."
    },
    {
     "name": "CompVis/EDGS",
     "url": "https://huggingface.co/spaces/CompVis/EDGS",
     "kind": "space",
     "note": "Demo space for the EDGS Gaussian splatting method."
    }
   ],
   "papers": [
    {
     "title": "High-Resolution Image Synthesis with Latent Diffusion Models",
     "url": "https://arxiv.org/abs/2112.10752",
     "year": "2022"
    },
    {
     "title": "Taming Transformers for High-Resolution Image Synthesis",
     "url": "https://arxiv.org/abs/2012.09841",
     "year": "2021"
    },
    {
     "title": "DepthFM: Fast Monocular Depth Estimation with Flow Matching",
     "url": "https://arxiv.org/abs/2403.13788",
     "year": "2024"
    },
    {
     "title": "CleanDIFT: Diffusion Features without Noise",
     "url": "https://arxiv.org/abs/2412.03439",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "latent-diffusion",
     "url": "https://github.com/CompVis/latent-diffusion",
     "note": "High-resolution image synthesis with latent diffusion"
    },
    {
     "name": "stable-diffusion",
     "url": "https://github.com/CompVis/stable-diffusion",
     "note": "Latent text-to-image diffusion model"
    },
    {
     "name": "taming-transformers",
     "url": "https://github.com/CompVis/taming-transformers",
     "note": "VQGAN and transformer image synthesis"
    },
    {
     "name": "stable-diffusion-v1-4",
     "url": "https://huggingface.co/CompVis/stable-diffusion-v1-4",
     "note": "Open weights on Hugging Face"
    }
   ],
   "openness": "Code is published on GitHub and weights on Hugging Face, with licences varying by project.",
   "key_people": [
    {
     "name": "Bjorn Ommer",
     "role": "Head of the Computer Vision and Learning group, LMU Munich",
     "url": "https://ommer-lab.com",
     "note": null
    },
    {
     "name": "Robin Rombach",
     "role": "First author of latent diffusion; former group PhD student",
     "url": "https://github.com/rromb",
     "note": "LDM first author, listed as a group PhD student on the lab site."
    },
    {
     "name": "Patrick Esser",
     "role": "Co-author of latent diffusion and Taming Transformers",
     "url": "https://github.com/pesser",
     "note": "Maintainer of CompVis repositories including stable-diffusion."
    },
    {
     "name": "Andreas Blattmann",
     "role": "Latent diffusion co-author; co-founder of Black Forest Labs",
     "url": "https://github.com/ablattmann",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The group's website lists its secretary, postal address and open positions on a contact page.",
   "sources": [
    "https://ommer-lab.com",
    "https://huggingface.co/CompVis",
    "https://github.com/CompVis",
    "https://arxiv.org/abs/2112.10752"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "compvis",
   "mlat": 48.11783,
   "mlon": 11.66166,
   "logo": true
  },
  {
   "orig_id": "contextual-ai",
   "name": "Contextual AI",
   "alias": null,
   "full_name": "Contextual AI",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "Mountain View",
   "lat": 37.3861,
   "lon": -122.0839,
   "website": "https://contextual.ai",
   "huggingface": "https://huggingface.co/ContextualAI",
   "github": "https://github.com/ContextualAI",
   "x": "https://x.com/contextualai",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Enterprise RAG company whose research arm releases open alignment libraries and retrieval and reranker models.",
   "tagline": "An enterprise retrieval-augmented generation company whose research arm releases open alignment methods and retrieval models.",
   "about": [
    "Contextual AI is a company founded in 2023 by Douwe Kiela and Amanpreet Singh, both formerly at Facebook AI Research and Hugging Face. It builds retrieval-augmented generation systems for enterprise use and pairs its commercial platform with research that publishes open code and models.",
    "The company raised a $20 million seed round led by Bain Capital Ventures in June 2023 and an $80 million Series A led by Greycroft in August 2024. Its research includes KTO, an alignment method that needs only a binary signal of whether an output is desirable, and GRIT, which trains one model for both generation and embedding.",
    "Its GitHub and Hugging Face organizations host alignment libraries, retrieval models and an instruction-following reranker released in March 2025, aimed at practitioners who build their own RAG pipelines."
   ],
   "mission": "The company works to make language models reliable on enterprise data by grounding them in retrieval. Its open releases are meant to share the underlying alignment and retrieval methods with the research community.",
   "notice": null,
   "known_for": [
    "Enterprise RAG platform",
    "Open alignment libraries",
    "Open reranker and retrieval models"
   ],
   "focus": [
    "RAG",
    "alignment",
    "embeddings",
    "rerankers"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Kiela and Singh founded the company to bring grounded language models to enterprises.",
     "year": "2023"
    },
    {
     "title": "KTO and GRIT papers",
     "note": "Published methods for preference-free alignment and for unified generative and embedding models, with released code.",
     "year": "2024"
    },
    {
     "title": "Series A",
     "note": "An $80 million round led by Greycroft.",
     "year": "2024"
    },
    {
     "title": "Instruction-following reranker",
     "note": "Released a reranker whose ranking can be steered with natural-language instructions.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "HALOs",
     "url": "https://github.com/ContextualAI/HALOs",
     "note": "Alignment methods library"
    },
    {
     "name": "gritlm",
     "url": "https://github.com/ContextualAI/gritlm",
     "note": "Generative representational instruction tuning"
    },
    {
     "name": "lens",
     "url": "https://github.com/ContextualAI/lens",
     "note": "LENS vision-language framework"
    }
   ],
   "hf_highlights": [
    {
     "name": "ctxl-rerank-v2-instruct-multilingual-2b",
     "url": "https://huggingface.co/ContextualAI/ctxl-rerank-v2-instruct-multilingual-2b",
     "kind": "model",
     "note": "2B reranker"
    },
    {
     "name": "ctxl-rerank-v2-instruct-multilingual-1b",
     "url": "https://huggingface.co/ContextualAI/ctxl-rerank-v2-instruct-multilingual-1b",
     "kind": "model",
     "note": "1B reranker"
    }
   ],
   "papers": [
    {
     "title": "KTO: Model Alignment as Prospect Theoretic Optimization",
     "url": "https://arxiv.org/abs/2402.01306",
     "year": "2024"
    },
    {
     "title": "Generative Representational Instruction Tuning",
     "url": "https://arxiv.org/abs/2402.09906",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "HALOs",
     "url": "https://github.com/ContextualAI/HALOs",
     "note": "Alignment methods library"
    },
    {
     "name": "gritlm",
     "url": "https://github.com/ContextualAI/gritlm",
     "note": "Generative representational instruction tuning"
    },
    {
     "name": "lens",
     "url": "https://github.com/ContextualAI/lens",
     "note": "LENS vision-language framework"
    }
   ],
   "openness": "Apache-2.0 and MIT code; open reranker and research models.",
   "key_people": [
    {
     "name": "Douwe Kiela",
     "role": "Co-founder; adjunct professor at Stanford",
     "url": "https://douwekiela.github.io",
     "note": "His site lists him as a former chief executive and now a researcher at Google DeepMind."
    },
    {
     "name": "Amanpreet Singh",
     "role": "Co-founder, former chief technology officer",
     "url": "https://apsdehal.in",
     "note": "His site describes the role as previous."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://contextual.ai",
    "https://github.com/ContextualAI",
    "https://huggingface.co/ContextualAI"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "contextual-ai",
   "mlat": 37.40775,
   "mlon": -122.0367,
   "logo": true
  },
  {
   "orig_id": "crane-ai-labs",
   "name": "Crane AI Labs",
   "alias": null,
   "full_name": "Crane AI Labs",
   "type": "startup",
   "founded": null,
   "country": "Uganda",
   "city": "Kampala",
   "lat": 0.3476,
   "lon": 32.5825,
   "website": null,
   "huggingface": "https://huggingface.co/CraneAILabs",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Crane AI Labs describes itself as building culturally relevant AI for the Global South. Its Hugging Face org hosts Swahili and Luganda Gemma-based models, tokenizers and benchmark datasets.",
   "tagline": "Ugandan AI group building open Luganda and Swahili language models, speech models and datasets on Gemma.",
   "about": [
    "Crane AI Labs is a Ugandan group co-founded by Kato Steven Mubiru and CTO Bakunga Bronson. Its Hugging Face organization hosts models, tokenizers and datasets for African languages.",
    "In 2025 it released Ganda Gemma 1B for Luganda and Swahili Gemma 1B, with GGUF and LiteRT builds. In 2026 it added Luganda education assistants, MedGemma-based health models for offline use by rural health workers and ASR models.",
    "Datasets such as UCCB, pedagogy benchmarks and a Luganda tokenizer evaluation support evaluation of low-resource languages."
   ],
   "mission": "The group describes its aim as culturally relevant AI for the Global South, delivered on small models that can run on-device.",
   "notice": null,
   "known_for": [
    "Ganda Gemma Luganda model",
    "Swahili Gemma 1B",
    "Edu-Ganda assistants",
    "African language datasets"
   ],
   "focus": [
    "language models",
    "low-resource languages",
    "datasets"
   ],
   "contributions": [
    {
     "title": "UCCB dataset",
     "note": "Released the UCCB dataset on Hugging Face.",
     "year": "2025"
    },
    {
     "title": "Ganda Gemma 1B",
     "note": "Released a Gemma-based 1B model for Luganda.",
     "year": "2025"
    },
    {
     "title": "Swahili Gemma 1B",
     "note": "Released a Swahili model with GGUF and LiteRT builds.",
     "year": "2025"
    },
    {
     "title": "Luganda tokenizer evaluation",
     "note": "Published tokenizer evaluation data.",
     "year": "2026"
    },
    {
     "title": "Crane ASR models",
     "note": "Released NeMo-based speech recognition models.",
     "year": "2026"
    },
    {
     "title": "Edu-Ganda Gemma",
     "note": "Released experimental Luganda primary-education assistants.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "ganda-gemma-1b-GGUF",
     "url": "https://huggingface.co/CraneAILabs/ganda-gemma-1b-GGUF",
     "kind": "model",
     "note": "Ganda Gemma."
    },
    {
     "name": "swahili-gemma-1b",
     "url": "https://huggingface.co/CraneAILabs/swahili-gemma-1b",
     "kind": "model",
     "note": "Swahili Gemma."
    },
    {
     "name": "UCCB",
     "url": "https://huggingface.co/datasets/CraneAILabs/UCCB",
     "kind": "dataset",
     "note": "UCCB."
    },
    {
     "name": "waxal-lug-clean",
     "url": "https://huggingface.co/datasets/CraneAILabs/waxal-lug-clean",
     "kind": "dataset",
     "note": "Luganda speech."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "swahili-gemma-1b",
     "url": "https://huggingface.co/CraneAILabs/swahili-gemma-1b",
     "note": "Swahili small model"
    },
    {
     "name": "ganda-gemma-1b GGUF",
     "url": "https://huggingface.co/CraneAILabs/ganda-gemma-1b-GGUF",
     "note": "Luganda small model"
    },
    {
     "name": "UCCB",
     "url": "https://huggingface.co/datasets/CraneAILabs/UCCB",
     "note": "Dataset"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face; base models carry their upstream licence terms. Country comes from press coverage, not the org page.",
   "key_people": [
    {
     "name": "Kato Steven Mubiru",
     "role": "Co-founder",
     "url": "https://deepmind.google/models/gemma/gemmaverse/craneai/",
     "note": "Listed on the Hugging Face organization team."
    },
    {
     "name": "Bakunga Bronson",
     "role": "Co-founder and chief technology officer",
     "url": "https://my.linkedin.com/in/bakungabronson",
     "note": "Also listed among engineers on Uganda's Project Crane."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website.",
   "sources": [
    "https://huggingface.co/CraneAILabs",
    "https://labscommunity.cohere.com/blog/2026/from-reviewing-words-to-building-africas-ai/"
   ],
   "confidence": "medium",
   "region": "Africa",
   "id": "crane-ai-labs",
   "logo": true
  },
  {
   "orig_id": "danish-foundation-models",
   "name": "Danish Foundation Models",
   "alias": null,
   "full_name": "Danish Foundation Models",
   "type": "collective",
   "founded": null,
   "country": "Denmark",
   "city": "Aarhus",
   "lat": 56.1629,
   "lon": 10.2039,
   "website": "https://foundationmodels.dk",
   "huggingface": "https://huggingface.co/danish-foundation-models",
   "github": "https://github.com/centre-for-humanities-computing",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/danish-foundation-models",
   "contact_email": null,
   "summary": "Danish community and research project building open foundation models, datasets and evaluations for Danish, led from Aarhus University's Center for Humanities Computing with partners.",
   "tagline": "A consortium of Danish universities building open language models, datasets and evaluations so that Danish is not left behind in AI.",
   "about": [
    "Danish Foundation Models is a collaborative effort across Danish universities and research institutions to develop, evaluate and adapt open language AI for Danish society. The core partners are the University of Southern Denmark, Aarhus University through its Center for Humanities Computing, the University of Copenhagen and the Alexandra Institute.",
    "Its members include professors Bolette Sandford Pedersen at Copenhagen, Kristoffer Nielbo at Aarhus and Peter Schneider-Kamp at Southern Denmark, with Kenneth Enevoldsen and many postdoctoral and PhD researchers. The project argues that smaller languages risk exclusion as language models become infrastructure, and it publishes models and datasets openly on Hugging Face.",
    "The 2023 paper Danish Foundation Models set out the program. Danish Dynaword, released in 2025, is an openly licensed, continuously updated collection of Danish text derived from the public parts of the Danish Gigaword, with a companion post-training dataset, dfm-dyna-instruct."
   ],
   "mission": "The consortium aims to give Danish society open, well-evaluated language technology that it can adopt and adapt. It treats openness of models and data as a condition for public benefit.",
   "notice": null,
   "known_for": [
    "Open Danish language models",
    "Danish Dynaword open text collection",
    "Scandinavian Embedding Benchmark",
    "Cross-university collaboration in Denmark"
   ],
   "focus": [
    "language models",
    "Danish",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Danish Foundation Models paper",
     "note": "The project's founding paper laid out open foundation models for Danish.",
     "year": "2023"
    },
    {
     "title": "Scandinavian Embedding Benchmarks",
     "note": "Aarhus-led benchmark for text embeddings across Scandinavian languages.",
     "year": "2024"
    },
    {
     "title": "Danish Dynaword",
     "note": "Openly licensed Danish text collection that is updated continuously.",
     "year": "2025"
    },
    {
     "title": "Dynaword paper",
     "note": "Preprint describing continuously developed open datasets.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "DaCy",
     "url": "https://github.com/centre-for-humanities-computing/DaCy",
     "note": "DaCy: The State of the Art Danish NLP pipeline using SpaCy (105 stars)"
    },
    {
     "name": "embedding-explorer",
     "url": "https://github.com/centre-for-humanities-computing/embedding-explorer",
     "note": "Tools for interactive visual exploration of semantic embeddings. (51 stars)"
    },
    {
     "name": "tweetopic",
     "url": "https://github.com/centre-for-humanities-computing/tweetopic",
     "note": "Blazing fast topic modelling for short texts. (37 stars)"
    },
    {
     "name": "odyCy",
     "url": "https://github.com/centre-for-humanities-computing/odyCy",
     "note": "A general-purpose NLP pipeline for Ancient Greek (29 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "danish-foundation-models/DFM-Mimir-v1.5",
     "url": "https://huggingface.co/danish-foundation-models/DFM-Mimir-v1.5",
     "kind": "model",
     "note": "Model: 3,684 downloads, 6 likes."
    },
    {
     "name": "danish-foundation-models/DFM-Mimir",
     "url": "https://huggingface.co/danish-foundation-models/DFM-Mimir",
     "kind": "model",
     "note": "Model: 3,033 downloads, 86 likes."
    },
    {
     "name": "danish-foundation-models/danish-dynaword",
     "url": "https://huggingface.co/datasets/danish-foundation-models/danish-dynaword",
     "kind": "dataset",
     "note": "Dataset: 16,676 downloads, 23 likes."
    },
    {
     "name": "danish-foundation-models/edda-danish-asr",
     "url": "https://huggingface.co/spaces/danish-foundation-models/edda-danish-asr",
     "kind": "space",
     "note": "Space: 2 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Danish Foundation Models",
     "url": "https://arxiv.org/abs/2311.07264",
     "year": "2023"
    },
    {
     "title": "The Scandinavian Embedding Benchmarks: Comprehensive Assessment of Multilingual and Monolingual Text Embedding",
     "url": "https://arxiv.org/abs/2406.02396",
     "year": "2024"
    },
    {
     "title": "Dynaword: From One-shot to Continuously Developed Datasets",
     "url": "https://arxiv.org/abs/2508.02271",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Danish Dynaword",
     "url": "https://huggingface.co/datasets/danish-foundation-models/danish-dynaword",
     "note": "Openly licensed Danish text corpus"
    },
    {
     "name": "Munin-7B-alpha",
     "url": "https://huggingface.co/danish-foundation-models/munin-7b-alpha",
     "note": "Danish-adapted language model"
    },
    {
     "name": "DaCy",
     "url": "https://github.com/centre-for-humanities-computing/DaCy",
     "note": "Danish NLP pipeline for spaCy"
    }
   ],
   "openness": "Datasets, models and code are public; the Dynaword corpus is built from openly licensed sources.",
   "key_people": [
    {
     "name": "Kristoffer Nielbo",
     "role": "Professor, Aarhus University",
     "url": "https://knielbo.github.io",
     "note": null
    },
    {
     "name": "Peter Schneider-Kamp",
     "role": "Professor, University of Southern Denmark",
     "url": "https://foundationmodels.dk/people.html",
     "note": null
    },
    {
     "name": "Bolette Sandford Pedersen",
     "role": "Professor, University of Copenhagen",
     "url": "https://foundationmodels.dk/people.html",
     "note": null
    },
    {
     "name": "Kenneth Enevoldsen",
     "role": "Assistant Professor, Aarhus University",
     "url": "https://foundationmodels.dk/people.html",
     "note": "First author of the Danish Foundation Models paper."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the project website.",
   "sources": [
    "https://foundationmodels.dk",
    "https://huggingface.co/danish-foundation-models",
    "https://github.com/centre-for-humanities-computing",
    "https://arxiv.org/abs/2311.07264"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "danish-foundation-models",
   "mlat": 56.18455,
   "mlon": 10.13655,
   "logo": true
  },
  {
   "orig_id": "deep-cogito",
   "name": "Deep Cogito",
   "alias": null,
   "full_name": "Deep Cogito",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.deepcogito.com",
   "huggingface": "https://huggingface.co/deepcogito",
   "github": "https://github.com/DeepCogito",
   "x": "https://x.com/deepcogito",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "AI lab releasing open-weight hybrid reasoning models trained with iterated distillation and amplification.",
   "tagline": "A San Francisco lab releasing open-weight reasoning models trained by iterated distillation and amplification, with the stated aim of general superintelligence.",
   "about": [
    "Deep Cogito was founded in San Francisco in 2024 by Drishan Arora, who led language-model work for generative search at Google, and Dhruv Malhotra, a former product lead at Google DeepMind. It describes its goal as building general superintelligence and argues that this requires advances in reasoning and iterative self-improvement.",
    "The lab introduced iterated distillation and amplification (IDA) with the Cogito v1 Preview models in April 2025, followed by Cogito v2 Preview in July 2025. Cogito v2.1, released in November 2025 as a 671 billion parameter open-weight model, was post-trained from an openly licensed DeepSeek base model.",
    "In August 2026 the company announced a $43 million Series A led by TQ Ventures, for total funding above $56 million. Reports say the money will fund hiring, training infrastructure, future open-weight Cogito models and an enterprise business."
   ],
   "mission": "Deep Cogito pursues models that improve through repeated cycles of search and distillation. It releases open weights as part of that research program.",
   "notice": null,
   "known_for": [
    "Iterated distillation and amplification",
    "Cogito open-weight reasoning models",
    "Cogito v2.1 release"
   ],
   "focus": [
    "language models",
    "post-training",
    "reasoning"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Arora and Malhotra started the lab in San Francisco.",
     "year": "2024"
    },
    {
     "title": "Cogito v1 Preview",
     "note": "The lab introduced IDA with its first preview models.",
     "year": "2025"
    },
    {
     "title": "Cogito v2 Preview",
     "note": "A second release moved from inference-time search toward self-improvement.",
     "year": "2025"
    },
    {
     "title": "Cogito v2.1",
     "note": "A 671 billion parameter open-weight model post-trained from an open DeepSeek base, released in November.",
     "year": "2025"
    },
    {
     "title": "Series A",
     "note": "A $43 million round led by TQ Ventures.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "cogito-v2",
     "url": "https://github.com/DeepCogito/cogito-v2",
     "note": "Cogito v2 model repo"
    }
   ],
   "hf_highlights": [
    {
     "name": "cogito-v1-preview-qwen-14B",
     "url": "https://huggingface.co/deepcogito/cogito-v1-preview-qwen-14B",
     "kind": "model",
     "note": "14B hybrid reasoning model"
    },
    {
     "name": "cogito-v2-preview-llama-109B-MoE",
     "url": "https://huggingface.co/deepcogito/cogito-v2-preview-llama-109B-MoE",
     "kind": "model",
     "note": "109B MoE model"
    },
    {
     "name": "cogito-v1-preview-llama-8B",
     "url": "https://huggingface.co/deepcogito/cogito-v1-preview-llama-8B",
     "kind": "model",
     "note": "8B model"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "cogito-v2",
     "url": "https://github.com/DeepCogito/cogito-v2",
     "note": "Cogito v2 model repo"
    }
   ],
   "openness": "Open model weights on Hugging Face (several under permissive or llama-style licences).",
   "key_people": [
    {
     "name": "Drishan Arora",
     "role": "Co-founder and chief executive",
     "url": "https://www.nasdaq.com/press-release/deep-cogito-raises-43m-series-advance-post-training-engine-frontier-intelligence-2026",
     "note": "Named in the company's Series A announcement; previously a senior engineer at Google."
    },
    {
     "name": "Dhruv Malhotra",
     "role": "Co-founder",
     "url": "https://www.nasdaq.com/press-release/deep-cogito-raises-43m-series-advance-post-training-engine-frontier-intelligence-2026",
     "note": "Previously a product lead at Google DeepMind; surname spelling varies across outlets."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.deepcogito.com",
    "https://huggingface.co/deepcogito",
    "https://github.com/DeepCogito"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "deep-cogito",
   "mlat": 37.79431,
   "mlon": -122.26958,
   "logo": true
  },
  {
   "orig_id": "deep-learning-indaba",
   "name": "Deep Learning Indaba",
   "alias": null,
   "full_name": "Deep Learning Indaba",
   "type": "nonprofit",
   "founded": 2017,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://deeplearningindaba.com",
   "huggingface": null,
   "github": "https://github.com/deep-learning-indaba",
   "x": "https://x.com/DeepIndaba",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/deep-learning-indaba/",
   "contact_email": "info@deeplearningindaba.com",
   "summary": "Deep Learning Indaba strengthens machine learning in Africa through an annual conference, schools and community programmes, run by a non-profit company registered in the UK. Its GitHub organisation publishes open practical notebooks from each Indaba.",
   "tagline": "A pan-African machine learning movement that strengthens African AI through an annual Indaba, local IndabaX events, mentorship and awards.",
   "about": [
    "Deep Learning Indaba was conceived to make Africans active shapers and owners of advances in artificial intelligence. The first Indaba ran from 10 to 15 September 2017 at the University of the Witwatersrand in Johannesburg with about 330 delegates, organized by a group that included Shakir Mohamed, Nyalleng Moorosi, Benjamin Rosman, Ulrich Paquet and Vukosi Marivate.",
    "The organization is led by Indaba Abantu, a steering group of Africans across the continent and abroad, with an advisory board and a board of trustees. Its legal entity is Deep Learning Indaba Ltd, a non-profit company limited by guarantee registered in England and Wales. Mohamed, a DeepMind research director, is listed as co-founder and chair of the trustees.",
    "Programs include the annual Indaba, locally organized IndabaX events, a mentorship program and awards. After Johannesburg in 2017 the gathering moved to Stellenbosch in 2018, and it later took place in Accra in 2023. The 2026 edition is planned for Nigeria, and the GitHub organization publishes open practical notebooks from each Indaba."
   ],
   "mission": "The Indaba aims to strengthen machine learning and artificial intelligence in Africa and to raise African participation and diversity in the field.",
   "notice": null,
   "known_for": [
    "Annual Indaba conference across Africa",
    "IndabaX local community events",
    "Mentorship and awards programmes",
    "Open practical notebooks"
   ],
   "focus": [
    "education",
    "community",
    "tooling"
   ],
   "contributions": [
    {
     "title": "First Indaba",
     "note": "Five-day series of introductory lectures in Johannesburg, hosted at Wits University with about 330 delegates.",
     "year": "2017"
    },
    {
     "title": "Indaba in Stellenbosch",
     "note": "The second annual gathering, hosted by Stellenbosch University from 9 to 14 September.",
     "year": "2018"
    },
    {
     "title": "Indaba in Accra",
     "note": "The 2023 edition was held in Ghana.",
     "year": "2023"
    },
    {
     "title": "2026 edition announced",
     "note": "The organization announced that the 2026 gathering would take place in Nigeria.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "indaba-pracs-2022",
     "url": "https://github.com/deep-learning-indaba/indaba-pracs-2022",
     "note": "2022 practicals."
    },
    {
     "name": "indaba-2018",
     "url": "https://github.com/deep-learning-indaba/indaba-2018",
     "note": "2018 practicals."
    },
    {
     "name": "indaba-pracs-2023",
     "url": "https://github.com/deep-learning-indaba/indaba-pracs-2023",
     "note": "2023 practicals."
    },
    {
     "name": "indaba-pracs-2019",
     "url": "https://github.com/deep-learning-indaba/indaba-pracs-2019",
     "note": "2019 practicals."
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Indaba practicals 2022",
     "url": "https://github.com/deep-learning-indaba/indaba-pracs-2022",
     "note": "Open notebooks for practicals"
    },
    {
     "name": "Indaba practicals 2023",
     "url": "https://github.com/deep-learning-indaba/indaba-pracs-2023",
     "note": "Open notebooks for practicals"
    },
    {
     "name": "Indaba 2018",
     "url": "https://github.com/deep-learning-indaba/indaba-2018",
     "note": "Practical notebooks"
    }
   ],
   "openness": "Teaching notebooks are open on GitHub; it does not train large models. Borderline inclusion as an education and community body.",
   "key_people": [
    {
     "name": "Shakir Mohamed",
     "role": "Co-founder and chair of the board of trustees",
     "url": "https://www.shakirm.com",
     "note": "Research director at DeepMind."
    },
    {
     "name": "Vukosi Marivate",
     "role": "Early organizer; professor at the University of Pretoria",
     "url": "https://www.vima.co.za",
     "note": "Chief investigator at Masakhane and founder of the Data Science for Social Impact group."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://deeplearningindaba.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://deeplearningindaba.com/about",
    "https://github.com/deep-learning-indaba"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "deep-learning-indaba",
   "mlat": 51.46616,
   "mlon": -0.04471,
   "logo": true
  },
  {
   "orig_id": "deepseek",
   "name": "DeepSeek",
   "alias": null,
   "full_name": "DeepSeek",
   "type": "startup",
   "founded": 2023,
   "country": "China",
   "city": "Hangzhou",
   "lat": 30.27,
   "lon": 120.16,
   "website": "https://www.deepseek.com",
   "huggingface": "https://huggingface.co/deepseek-ai",
   "github": "https://github.com/deepseek-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "service@deepseek.com",
   "summary": "Hangzhou-based AI lab that publishes large language and reasoning models. Its Hugging Face organisation hosts more than 100 models and is among the most followed in the field.",
   "tagline": "A Hangzhou AI lab founded by Liang Wenfeng and owned by the hedge fund High-Flyer, known for open-weight V3, R1 and V4 models released under the MIT License.",
   "about": [
    "DeepSeek was founded on 17 July 2023 as a spin-off from the AI research lab of High-Flyer, a Chinese hedge fund that owns and backs it. Liang Wenfeng founded the company and serves as CEO, and it had about 160 employees in 2025.",
    "The lab released DeepSeek-V2 in May 2024, which introduced multi-head latent attention within a mixture-of-experts design, and DeepSeek-V3 in December 2024. DeepSeek-R1 followed on 20 January 2025 alongside a chatbot, with weights under the MIT License. Later releases include V3.1 in August 2025, V3.2 in December 2025 and a V4 preview in April 2026.",
    "Wikipedia reports that DeepSeek closed a $7 billion Series A at a $52 billion post-money valuation in May 2026, ending its reliance on its parent alone. The company publishes weights and technical reports but not training data, and several governments have restricted its use on official systems."
   ],
   "mission": "DeepSeek pursues capable general models and publishes weights and technical reports for them, funded originally by High-Flyer and, from 2026, by outside investors as well.",
   "notice": null,
   "known_for": [
    "DeepSeek-R1 reasoning model",
    "MIT-licensed open weights",
    "Multi-head latent attention",
    "DeepSeek-V3 mixture-of-experts model"
   ],
   "focus": [
    "language models",
    "reasoning",
    "Chinese",
    "English"
   ],
   "contributions": [
    {
     "title": "DeepSeek founded",
     "note": "The company was established as a spin-off of High-Flyer's AI research lab.",
     "year": "2023"
    },
    {
     "title": "DeepSeek-V2",
     "note": "The model introduced multi-head latent attention with a mixture-of-experts architecture.",
     "year": "2024"
    },
    {
     "title": "DeepSeek-V3",
     "note": "A larger mixture-of-experts model was released in December 2024.",
     "year": "2024"
    },
    {
     "title": "DeepSeek-R1",
     "note": "The reasoning model launched on 20 January 2025 and was released under the MIT License.",
     "year": "2025"
    },
    {
     "title": "DeepSeek-V3.2",
     "note": "Released in December 2025, following V3.1 in August 2025.",
     "year": "2025"
    },
    {
     "title": "DeepSeek-V4 preview",
     "note": "Previewed in April 2026, with Flash and Pro versions released later that year.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "DeepSeek-V3",
     "url": "https://github.com/deepseek-ai/DeepSeek-V3",
     "note": "Repository for the V3 model, one of the most starred in the org."
    },
    {
     "name": "DeepSeek-R1",
     "url": "https://github.com/deepseek-ai/DeepSeek-R1",
     "note": "Repository for the R1 reasoning model."
    },
    {
     "name": "awesome-deepseek-integration",
     "url": "https://github.com/deepseek-ai/awesome-deepseek-integration",
     "note": "Community list of software that integrates the DeepSeek API."
    },
    {
     "name": "deepseek-harness",
     "url": "https://github.com/deepseek-ai/deepseek-harness",
     "note": "Plugin-style agent harness, the org's most starred repo."
    }
   ],
   "hf_highlights": [
    {
     "name": "DeepSeek-V3.2",
     "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2",
     "kind": "model",
     "note": "Most downloaded model in the organisation."
    },
    {
     "name": "DeepSeek-V4-Flash-0731",
     "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731",
     "kind": "model",
     "note": "Recent V4 Flash release."
    },
    {
     "name": "DeepSeek-OCR",
     "url": "https://huggingface.co/deepseek-ai/DeepSeek-OCR",
     "kind": "model",
     "note": "Open OCR model."
    },
    {
     "name": "DeepSeek-ProverBench",
     "url": "https://huggingface.co/datasets/deepseek-ai/DeepSeek-ProverBench",
     "kind": "dataset",
     "note": "Benchmark for formal theorem proving."
    }
   ],
   "papers": [
    {
     "title": "DeepSeek-V3 Technical Report",
     "url": "https://arxiv.org/abs/2412.19437",
     "year": "2024"
    },
    {
     "title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning",
     "url": "https://arxiv.org/abs/2501.12948",
     "year": "2025"
    },
    {
     "title": "DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model",
     "url": "https://arxiv.org/abs/2405.04434",
     "year": "2024"
    },
    {
     "title": "DeepSeek LLM: Scaling Open-Source Language Models with Longtermism",
     "url": "https://arxiv.org/abs/2401.02954",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "DeepSeek on Hugging Face",
     "url": "https://huggingface.co/deepseek-ai",
     "note": "106 models, 30 papers"
    }
   ],
   "openness": "Model weights are published on Hugging Face; licences vary by release (check each model card).",
   "key_people": [
    {
     "name": "Liang Wenfeng",
     "role": "Founder and CEO",
     "url": "https://en.wikipedia.org/wiki/Liang_Wenfeng",
     "note": "Also co-founder of the parent fund High-Flyer."
    }
   ],
   "discord_note": null,
   "contact_note": "The official site prints a service email address; code and model questions go through GitHub issues and Hugging Face discussions.",
   "sources": [
    "https://huggingface.co/deepseek-ai",
    "https://www.deepseek.com",
    "https://huggingface.co/api/organizations/deepseek-ai/overview"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "deepseek",
   "logo": true
  },
  {
   "orig_id": "deepset",
   "name": "deepset",
   "alias": null,
   "full_name": "deepset",
   "type": "startup",
   "founded": 2018,
   "country": "Germany",
   "city": "Berlin",
   "lat": 52.52,
   "lon": 13.405,
   "website": "https://www.deepset.ai",
   "huggingface": "https://huggingface.co/deepset",
   "github": "https://github.com/deepset-ai",
   "x": "https://x.com/deepset_ai",
   "discord": "https://discord.com/invite/Dr63fr9NDS",
   "linkedin": "https://www.linkedin.com/company/deepset-ai",
   "contact_email": null,
   "summary": "Berlin company that maintains Haystack, an open-source framework for building LLM and RAG applications, and has released German language models and datasets.",
   "tagline": "A Berlin company that created and maintains Haystack, an open-source framework for building LLM and retrieval-augmented applications.",
   "about": [
    "deepset began building custom language AI systems for industry in 2018. In 2020 it developed Haystack, an open-source Python framework that has become a widely used base for AI agents and retrieval applications.",
    "The company is co-led by Milos Rusic as CEO and Malte Pietsch as CTO, both co-founders. It sells enterprise products and services built on the open-source foundation and has also released German language models and datasets."
   ],
   "mission": "deepset aims to make custom AI solutions accessible to every organization by simplifying the development of LLM agents and applications.",
   "notice": null,
   "known_for": [
    "Haystack open-source framework",
    "Retrieval-augmented generation tooling",
    "German language models and datasets"
   ],
   "focus": [
    "RAG",
    "tooling",
    "German",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Custom language AI",
     "note": "The company began building language AI systems for industry customers.",
     "year": "2018"
    },
    {
     "title": "Haystack",
     "note": "deepset developed Haystack as an open-source Python framework.",
     "year": "2020"
    }
   ],
   "github_highlights": [
    {
     "name": "haystack",
     "url": "https://github.com/deepset-ai/haystack",
     "note": "Open-source AI orchestration framework for building context-engineered, production-ready LLM applications. Des (27,022 stars)"
    },
    {
     "name": "FARM",
     "url": "https://github.com/deepset-ai/FARM",
     "note": ":house_with_garden: Fast & easy transfer learning for NLP. Harvesting language models for the industry. Focus  (1,755 stars)"
    },
    {
     "name": "haystack-cookbook",
     "url": "https://github.com/deepset-ai/haystack-cookbook",
     "note": "👩🏻‍🍳 A collection of example notebooks using Haystack (551 stars)"
    },
    {
     "name": "haystack-tutorials",
     "url": "https://github.com/deepset-ai/haystack-tutorials",
     "note": "Here you can find all the Tutorials for Haystack 📓 (378 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "deepset/roberta-base-squad2",
     "url": "https://huggingface.co/deepset/roberta-base-squad2",
     "kind": "model",
     "note": "Model: 351,349 downloads, 979 likes."
    },
    {
     "name": "deepset/bert-large-uncased-whole-word-masking-squad2",
     "url": "https://huggingface.co/deepset/bert-large-uncased-whole-word-masking-squad2",
     "kind": "model",
     "note": "Model: 201,585 downloads, 31 likes."
    },
    {
     "name": "deepset/prompt-injections",
     "url": "https://huggingface.co/datasets/deepset/prompt-injections",
     "kind": "dataset",
     "note": "Dataset: 15,313 downloads, 200 likes."
    },
    {
     "name": "deepset/autoquizzer",
     "url": "https://huggingface.co/spaces/deepset/autoquizzer",
     "kind": "space",
     "note": "Space: 226 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "GermanQuAD and GermanDPR: Improving Non-English Question Answering and Passage Retrieval",
     "url": "https://arxiv.org/abs/2104.12741",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Haystack",
     "url": "https://github.com/deepset-ai/haystack",
     "note": "Open-source AI orchestration framework"
    },
    {
     "name": "roberta-base-squad2",
     "url": "https://huggingface.co/deepset/roberta-base-squad2",
     "note": "Question answering model"
    },
    {
     "name": "gbert-large",
     "url": "https://huggingface.co/deepset/gbert-large",
     "note": "German BERT model"
    },
    {
     "name": "prompt-injections",
     "url": "https://huggingface.co/datasets/deepset/prompt-injections",
     "note": "Prompt-injection dataset"
    }
   ],
   "openness": "Haystack is Apache-2.0 and most models are openly licensed; the commercial platform is separate.",
   "key_people": [
    {
     "name": "Milos Rusic",
     "role": "CEO and co-founder",
     "url": "https://www.deepset.ai/about",
     "note": null
    },
    {
     "name": "Malte Pietsch",
     "role": "CTO and co-founder",
     "url": "https://www.deepset.ai/about",
     "note": null
    },
    {
     "name": "Ole Dallerup",
     "role": "VP Engineering and Product",
     "url": "https://www.deepset.ai/about",
     "note": null
    }
   ],
   "discord_note": "Community Discord for Haystack users and contributors.",
   "contact_note": "Contact through the website and Discord.",
   "sources": [
    "https://www.deepset.ai",
    "https://huggingface.co/deepset",
    "https://github.com/deepset-ai"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "deepset",
   "mlat": 52.4767,
   "mlon": 13.405,
   "logo": true
  },
  {
   "orig_id": "dfki",
   "name": "DFKI",
   "alias": null,
   "full_name": "DFKI",
   "type": "institute",
   "founded": 1988,
   "country": "Germany",
   "city": "Kaiserslautern",
   "lat": 49.4401,
   "lon": 7.7491,
   "website": "https://www.dfki.de",
   "huggingface": "https://huggingface.co/DFKI",
   "github": "https://github.com/DFKI",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "DFKI is the German Research Center for Artificial Intelligence, a non-profit research company with several sites, publishing research code and datasets on GitHub and Hugging Face.",
   "tagline": "DFKI is Germany's national AI research center, a non-profit public-private partnership founded in 1988 that turns academic research into industrial applications.",
   "about": [
    "The German Research Center for Artificial Intelligence was established in 1988 as a non-profit public-private partnership. It began in Kaiserslautern and Saarbrücken, and Wolfgang Wahlster has served as its scientific director since the founding.",
    "Today DFKI operates research facilities in Kaiserslautern, Saarbrücken, Bremen, Oldenburg and Osnabrück, laboratories in Berlin, Darmstadt and Lübeck, and a branch office in Trier. Its work is organized into 29 research departments, thirteen competence centers and eight living labs, and it is funded by the European Union, federal ministries, the German states and the German Research Foundation.",
    "DFKI led the national Verbmobil project on speech translation, which produced several spin-off companies and earned the 2001 German Future Prize. Its public code and datasets appear on GitHub and Hugging Face, where researchers in several departments publish benchmarks and models."
   ],
   "mission": "To conduct application-oriented AI research for industry and society and to transfer results into commercial software and spin-offs.",
   "notice": "DFKI has industry shareholders and releases some, not all, of its software as open source.",
   "known_for": [
    "National AI research institute of Germany",
    "Verbmobil speech-translation project",
    "Technology transfer and spin-off companies"
   ],
   "focus": [
    "AI research"
   ],
   "contributions": [
    {
     "title": "Founded in Kaiserslautern and Saarbrücken",
     "note": "DFKI was established as a non-profit public-private partnership for applied AI research.",
     "year": "1988"
    },
    {
     "title": "Verbmobil recognised with the German Future Prize",
     "note": "The national speech-translation project that DFKI led, translating spontaneous German, English and Japanese speech, won the award for its leader's group.",
     "year": "2001"
    },
    {
     "title": "Public GitHub organization opened",
     "note": "DFKI created its GitHub organization to publish research code, later adding the Awesome-DFKI index of its public projects.",
     "year": "2017"
    },
    {
     "title": "CGHD handwritten circuit dataset",
     "note": "DFKI researchers released a public ground-truth dataset of handwritten circuit diagram images.",
     "year": "2022"
    },
    {
     "title": "SphereCraft paper repository",
     "note": "The most-starred repository in the DFKI GitHub organization accompanies the SphereCraft paper.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "spherecrafthub",
     "url": "https://github.com/DFKI/spherecrafthub",
     "note": "The Github page for the SphereCraft paper (32 stars)"
    },
    {
     "name": "Awesome-DFKI",
     "url": "https://github.com/DFKI/Awesome-DFKI",
     "note": "A curated list of awesome DFKI GitLabs, Groups and Projects. (27 stars)"
    },
    {
     "name": "treesatai_benchmark",
     "url": "https://github.com/DFKI/treesatai_benchmark",
     "note": "Tree species classification using LightGBM in the context of project TreeSatAI (13 stars)"
    },
    {
     "name": "cghd",
     "url": "https://github.com/DFKI/cghd",
     "note": "A Public Ground-Truth Dataset for Handwritten Circuit Diagram Images (12 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "DFKI/glove-th-cc100",
     "url": "https://huggingface.co/DFKI/glove-th-cc100",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "DFKI/glove-sq-cc100",
     "url": "https://huggingface.co/DFKI/glove-sq-cc100",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "DFKI/radr_intents",
     "url": "https://huggingface.co/datasets/DFKI/radr_intents",
     "kind": "dataset",
     "note": "Dataset with 0 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "SphereCraft",
     "url": "https://github.com/DFKI/spherecrafthub",
     "note": "Repository for the SphereCraft paper"
    },
    {
     "name": "Awesome-DFKI",
     "url": "https://github.com/DFKI/Awesome-DFKI",
     "note": "List of DFKI open projects"
    },
    {
     "name": "TreeSatAI benchmark",
     "url": "https://github.com/DFKI/treesatai_benchmark",
     "note": "Tree species classification in remote sensing"
    },
    {
     "name": "cghd",
     "url": "https://github.com/DFKI/cghd",
     "note": "Dataset of handwritten circuit diagram images"
    }
   ],
   "openness": "Multiple public repositories and Hugging Face datasets; DFKI is broad, so openness varies by project.",
   "key_people": [
    {
     "name": "Antonio Krüger",
     "role": "Chief Executive Officer",
     "url": "https://www.dfki.de/en/web/about-us",
     "note": null
    },
    {
     "name": "Wolfgang Wahlster",
     "role": "Scientific director since 1988, led Verbmobil",
     "url": "https://www.dfki.de/~wahlster/Curriculum_Vitae_Wahlster.html",
     "note": null
    },
    {
     "name": "Patrick Schramowski",
     "role": "Researcher, DFKI Hugging Face organization",
     "url": "https://huggingface.co/PSaiml",
     "note": null
    },
    {
     "name": "Manuel Brack",
     "role": "Researcher, DFKI Hugging Face organization",
     "url": "https://huggingface.co/mbrack",
     "note": null
    },
    {
     "name": "Simon Ostermann",
     "role": "Researcher, DFKI Hugging Face organization",
     "url": "https://huggingface.co/simonost",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.dfki.de",
    "https://huggingface.co/DFKI",
    "https://github.com/DFKI"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "dfki",
   "logo": true
  },
  {
   "orig_id": "dicta",
   "name": "DICTA: Israel Center for Text Analysis",
   "alias": null,
   "full_name": "DICTA: Israel Center for Text Analysis",
   "type": "nonprofit",
   "founded": null,
   "country": "Israel",
   "city": "Jerusalem",
   "lat": 31.77,
   "lon": 35.22,
   "website": "https://dicta.org.il/",
   "huggingface": "https://huggingface.co/dicta-il",
   "github": "https://github.com/Dicta-Israel-Center-for-Text-Analysis",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Non-profit building Hebrew language models and tools, including the DictaLM open-weight collection.",
   "tagline": "A non-profit center building Hebrew language models and text tools, including the open DictaLM collection.",
   "about": [
    "DICTA, the Israel Center for Text Analysis, was founded by Professor Moshe Koppel of Bar-Ilan University and builds language technology for Hebrew. Its early tools served Jewish studies and digital humanities researchers, including the Nakdan vowelization tool, a Scriptural Allusion Finder and an Abbreviation Expander.",
    "The core research team includes Shaltiel Shmidman, Avi Shmidman, who holds dual affiliations with DICTA and Bar-Ilan University, and Koppel. Its Hugging Face organization dicta-il and its GitHub organization publish models and tools for Hebrew.",
    "Its model line began with DictaBERT in 2023 and DictaLM in the same year. DictaLM 2.0 followed in 2024 with an enlarged Hebrew vocabulary, and the DictaLM 3.0 report of February 2026 presents open 24B, 12B and 1.7B models adapted from Mistral Small 3.1, NVIDIA Nemotron Nano V2 and Qwen3, hosted under Apache 2.0."
   ],
   "mission": "The center works to provide high-quality, openly available language technology for Hebrew.",
   "notice": null,
   "known_for": [
    "Hebrew language models",
    "DictaLM open-weight collection",
    "Hebrew text analysis tools"
   ],
   "focus": [
    "language models",
    "Hebrew",
    "NLP tools"
   ],
   "contributions": [
    {
     "title": "Nakdan",
     "note": "Presented a professional Hebrew diacritizer at the ACL system demonstrations.",
     "year": "2020"
    },
    {
     "title": "Extra-large vocabulary Hebrew BERT",
     "note": "Published a contrastive analysis of Hebrew BERT models and a new model.",
     "year": "2022"
    },
    {
     "title": "DictaBERT and DictaLM",
     "note": "Released a BERT suite and the first generative DictaLM for Modern Hebrew.",
     "year": "2023"
    },
    {
     "title": "DictaLM 2.0",
     "note": "Adapted LLMs to Hebrew with enhanced vocabulary and instruction capabilities.",
     "year": "2024"
    },
    {
     "title": "DictaLM 3.0",
     "note": "Open Hebrew-centric 24B, 12B and 1.7B models with base and chat variants.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "alephbertgimmel",
     "url": "https://github.com/Dicta-Israel-Center-for-Text-Analysis/alephbertgimmel",
     "note": "Hebrew BERT with 128K vocabulary."
    },
    {
     "name": "Talmud-Bavli-with-Nikud",
     "url": "https://github.com/Dicta-Israel-Center-for-Text-Analysis/Talmud-Bavli-with-Nikud",
     "note": "Talmud text with vowel marks."
    },
    {
     "name": "Dicta-Library-Download",
     "url": "https://github.com/Dicta-Israel-Center-for-Text-Analysis/Dicta-Library-Download",
     "note": "Library download tool."
    }
   ],
   "hf_highlights": [
    {
     "name": "dicta-il/dictabert-joint",
     "url": "https://huggingface.co/dicta-il/dictabert-joint",
     "kind": "model",
     "note": "Joint Hebrew analysis model."
    },
    {
     "name": "dicta-il/DictaLM-3.0-Nemotron-12B-Instruct-GGUF",
     "url": "https://huggingface.co/dicta-il/DictaLM-3.0-Nemotron-12B-Instruct-GGUF",
     "kind": "model",
     "note": "Hebrew LLM."
    },
    {
     "name": "dicta-il/dictabert",
     "url": "https://huggingface.co/dicta-il/dictabert",
     "kind": "model",
     "note": "Hebrew BERT."
    },
    {
     "name": "dicta-il/dictabert-ner",
     "url": "https://huggingface.co/dicta-il/dictabert-ner",
     "kind": "model",
     "note": "Hebrew NER."
    }
   ],
   "papers": [
    {
     "title": "DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew",
     "url": "https://arxiv.org/abs/2308.16687",
     "year": "2023"
    },
    {
     "title": "Introducing DictaLM -- A Large Generative Language Model for Modern Hebrew",
     "url": "https://arxiv.org/abs/2309.14568",
     "year": "2023"
    },
    {
     "title": "Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction Capabilities",
     "url": "https://arxiv.org/abs/2407.07080",
     "year": "2024"
    },
    {
     "title": "Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs",
     "url": "https://arxiv.org/abs/2602.02104",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "dictabert-lex",
     "url": "https://huggingface.co/dicta-il/dictabert-lex",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "dictabert-joint",
     "url": "https://huggingface.co/dicta-il/dictabert-joint",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open-weight Hebrew models.",
   "key_people": [
    {
     "name": "Moshe Koppel",
     "role": "Founder; professor of computer science, Bar-Ilan University",
     "url": "https://u.cs.biu.ac.il/~koppel/",
     "note": null
    },
    {
     "name": "Avi Shmidman",
     "role": "Researcher, DICTA and Bar-Ilan University",
     "url": "https://dicta.org.il/",
     "note": "Lead author on DictaBERT and DictaLM papers."
    },
    {
     "name": "Shaltiel Shmidman",
     "role": "Researcher, DICTA",
     "url": "https://dicta.org.il/",
     "note": "First author of DictaBERT and DictaLM 2.0."
    }
   ],
   "discord_note": null,
   "contact_note": "dicta.org.il hosts the tools and contact details.",
   "sources": [
    "https://huggingface.co/dicta-il",
    "https://dicta.org.il/"
   ],
   "confidence": "high",
   "region": "Middle East",
   "id": "dicta-israel-center-for-text-analysis",
   "logo": true
  },
  {
   "orig_id": "digital-umuganda",
   "name": "Digital Umuganda",
   "alias": null,
   "full_name": "Digital Umuganda",
   "type": "nonprofit",
   "founded": null,
   "country": "Rwanda",
   "city": "Kigali",
   "lat": -1.9441,
   "lon": 30.0619,
   "website": null,
   "huggingface": "https://huggingface.co/DigitalUmuganda",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Digital Umuganda works on African language NLP, speech and data collection, with a focus on Kinyarwanda and other languages. Its Hugging Face org hosts speech models and the AfriVoice datasets.",
   "tagline": "A Rwanda-based team building voice infrastructure and open datasets for African languages, with a focus on Kinyarwanda.",
   "about": [
    "Digital Umuganda is a Rwanda-based African startup whose stated goal is to create and promote voice infrastructure for African languages. Its interests are African natural language processing, speech-to-text, text-to-speech and data collection.",
    "Its Hugging Face organisation lists 12 members, 39 models and 22 datasets. Releases include Kinyarwanda and Swahili speech recognition models, a Kinyarwanda translation model, a text-to-speech model and the AfriVoice speech datasets."
   ],
   "mission": "Digital Umuganda aims to build the speech and language resources that African languages need to be supported by modern AI.",
   "notice": null,
   "known_for": [
    "Kinyarwanda speech recognition",
    "AfriVoice speech datasets",
    "African language data collection"
   ],
   "focus": [
    "speech",
    "datasets",
    "low-resource languages"
   ],
   "contributions": [
    {
     "title": "Kinyarwanda voice hackathon",
     "note": "A 2019 Mozilla and GIZ hackathon in Kigali launched the Kinyarwanda voice data effort; the winning idea adapted the umuganda community-work tradition to collect speech.",
     "year": "2019"
    },
    {
     "title": "Common Voice contributions",
     "note": "The company became a leading African contributor to Mozilla Common Voice, building open Kinyarwanda speech data and models.",
     "year": "2019-2022"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "whisper_small_kinyarwanda",
     "url": "https://huggingface.co/DigitalUmuganda/whisper_small_kinyarwanda",
     "kind": "model",
     "note": "Kinyarwanda ASR."
    },
    {
     "name": "lingala_vits_tts",
     "url": "https://huggingface.co/DigitalUmuganda/lingala_vits_tts",
     "kind": "model",
     "note": "Lingala TTS."
    },
    {
     "name": "Afrivoice",
     "url": "https://huggingface.co/datasets/DigitalUmuganda/Afrivoice",
     "kind": "dataset",
     "note": "Afrivoice."
    },
    {
     "name": "Afrivoice_Swahili",
     "url": "https://huggingface.co/datasets/DigitalUmuganda/Afrivoice_Swahili",
     "kind": "dataset",
     "note": "Swahili speech."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "AfriVoice",
     "url": "https://huggingface.co/datasets/DigitalUmuganda/Afrivoice",
     "note": "Speech dataset collection"
    },
    {
     "name": "Whisper small Kinyarwanda",
     "url": "https://huggingface.co/DigitalUmuganda/whisper_small_kinyarwanda",
     "note": "ASR model"
    },
    {
     "name": "Kinyarwanda TTS",
     "url": "https://huggingface.co/DigitalUmuganda/KinyarwandaTTS_female_voice",
     "note": "TTS voice model"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face; licences per repo. Only the Hugging Face org was opened, so website and legal form are unverified.",
   "key_people": [
    {
     "name": "Audace Niyonkuru",
     "role": "Founder and CEO",
     "url": "https://blogs.nvidia.com/blog/common-voice-conversational-ai",
     "note": "Named as founder and CEO in NVIDIA and Mozilla coverage of the Common Voice work."
    },
    {
     "name": "Ali Nyiringabo",
     "role": "Co-founder",
     "url": "https://blogs.nvidia.com/blog/common-voice-conversational-ai",
     "note": "Named as a co-founder in NVIDIA coverage."
    },
    {
     "name": "Samuel Rutunda",
     "role": "Team member, Hugging Face organization",
     "url": "https://huggingface.co/rutsam",
     "note": "Listed among the twelve members of the Hugging Face organization."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website.",
   "sources": [
    "https://huggingface.co/DigitalUmuganda"
   ],
   "confidence": "medium",
   "region": "Africa",
   "id": "digital-umuganda",
   "logo": true
  },
  {
   "orig_id": "dtai-kuleuven",
   "name": "DTAI",
   "alias": "KU Leuven",
   "full_name": "DTAI (KU Leuven)",
   "type": "academic",
   "founded": null,
   "country": "Belgium",
   "city": "Leuven",
   "lat": 50.8798,
   "lon": 4.7005,
   "website": "https://dtai.cs.kuleuven.be",
   "huggingface": "https://huggingface.co/DTAI-KULeuven",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Machine learning and AI group at KU Leuven; its researchers released the RobBERT Dutch language models.",
   "tagline": "The DTAI group at KU Leuven combines machine learning, logic and data mining and is the home of the RobBERT Dutch language models.",
   "about": [
    "DTAI is the Declarative Languages and Artificial Intelligence research group in the Department of Computer Science at KU Leuven in Belgium. Its academic staff includes Hendrik Blockeel, Jesse Davis, Luc De Raedt, Tias Guns and Tom Schrijvers, and it is linked to the ADVISE, EAVISE and M-Group research groups.",
    "In January 2020 researchers Pieter Delobelle, Thomas Winters and Bettina Berendt introduced RobBERT, a Dutch RoBERTa-based language model. The group has since refreshed it with RobBERT-2022 and RobBERT-2023, which are published under the DTAI-KULeuven organization on Hugging Face under the MIT license."
   ],
   "mission": "To advance machine learning and reasoning methods and to supply open language resources for Dutch.",
   "notice": null,
   "known_for": [
    "RobBERT Dutch language models",
    "Machine learning combined with logic and constraints",
    "KU Leuven AI research tradition"
   ],
   "focus": [
    "Dutch NLP",
    "machine learning"
   ],
   "contributions": [
    {
     "title": "RobBERT released",
     "note": "A Dutch RoBERTa-based language model was introduced in a paper by Delobelle, Winters and Berendt.",
     "year": "2020"
    },
    {
     "title": "RobBERT-2022",
     "note": "An update with an extended tokenizer and further pretraining on recent Dutch OSCAR data to reflect evolving language use.",
     "year": "2022"
    },
    {
     "title": "RobBERT-2023 models published",
     "note": "A Dutch base model initialized from RoBERTa weights and a first 355M-parameter large Dutch variant were released under the MIT license.",
     "year": "2023"
    },
    {
     "title": "RobBERT-2023 journal paper",
     "note": "Delobelle and Remy published the model conversion method in the Computational Linguistics in the Netherlands Journal.",
     "year": "2024"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "DTAI-KULeuven/robbert-2023-dutch-large",
     "url": "https://huggingface.co/DTAI-KULeuven/robbert-2023-dutch-large",
     "kind": "model",
     "note": "Model with 23 likes on the Hub."
    },
    {
     "name": "DTAI-KULeuven/robbert-v2-dutch-sentiment",
     "url": "https://huggingface.co/DTAI-KULeuven/robbert-v2-dutch-sentiment",
     "kind": "model",
     "note": "Model with 9 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "RobBERT: a Dutch RoBERTa-based Language Model",
     "url": "https://arxiv.org/abs/2001.06286",
     "year": "2020"
    },
    {
     "title": "RobBERT-2022: Updating a Dutch Language Model to Account for Evolving Language Use",
     "url": "https://arxiv.org/abs/2211.08192",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "RobBERT-2023 large",
     "url": "https://huggingface.co/DTAI-KULeuven/robbert-2023-dutch-large",
     "note": "Dutch RoBERTa-based model"
    },
    {
     "name": "RobBERT 2022 base",
     "url": "https://huggingface.co/DTAI-KULeuven/robbert-2022-dutch-base",
     "note": "Dutch base model"
    },
    {
     "name": "RobBERT v2 sentiment",
     "url": "https://huggingface.co/DTAI-KULeuven/robbert-v2-dutch-sentiment",
     "note": "Dutch sentiment model"
    }
   ],
   "openness": "RobBERT models are public on Hugging Face under the DTAI-KULeuven organisation.",
   "key_people": [
    {
     "name": "Luc De Raedt",
     "role": "Professor, DTAI academic staff",
     "url": "http://www.kuleuven.be/wieiswie/en/person/u0010316",
     "note": null
    },
    {
     "name": "Hendrik Blockeel",
     "role": "Professor, DTAI academic staff",
     "url": "http://www.kuleuven.be/wieiswie/en/person/u0008122",
     "note": null
    },
    {
     "name": "Jesse Davis",
     "role": "Professor, DTAI academic staff",
     "url": "http://people.cs.kuleuven.be/~jesse.davis/",
     "note": null
    },
    {
     "name": "Tias Guns",
     "role": "Professor, DTAI academic staff",
     "url": "https://www.kuleuven.be/wieiswie/nl/person/00056117",
     "note": null
    },
    {
     "name": "Pieter Delobelle",
     "role": "Lead author of RobBERT",
     "url": "https://huggingface.co/pdelobelle",
     "note": null
    },
    {
     "name": "Thomas Winters",
     "role": "RobBERT co-author",
     "url": "https://huggingface.co/thomaswint",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://dtai.cs.kuleuven.be",
    "https://huggingface.co/DTAI-KULeuven",
    "https://arxiv.org/abs/2001.06286"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "dtai",
   "logo": true
  },
  {
   "orig_id": "edinburghnlp",
   "name": "EdinburghNLP",
   "alias": null,
   "full_name": "EdinburghNLP",
   "type": "academic",
   "founded": null,
   "country": "United Kingdom",
   "city": "Edinburgh",
   "lat": 55.9533,
   "lon": -3.1883,
   "website": "https://edinburghnlp.inf.ed.ac.uk",
   "huggingface": "https://huggingface.co/EdinburghNLP",
   "github": "https://github.com/EdinburghNLP",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Natural language processing group at the University of Edinburgh, behind open tools and datasets including the Nematus NMT toolkit and the XSum summarisation dataset.",
   "tagline": "EdinburghNLP is the University of Edinburgh's natural language processing group, ranked among the most productive in the world and the home of Nematus.",
   "about": [
    "EdinburghNLP brings together faculty, postdocs and PhD students in the School of Informatics. Its research covers language modelling, morphology, parsing, semantics, discourse, generation and machine translation, with links to speech, vision and robotics.",
    "The group reports a faculty of more than a dozen core members and five European Research Council grants, and it is ranked first for NLP output by csrankings.org. It publishes code and data openly, including the Nematus neural machine translation toolkit, the XSum summarization dataset and a widely used list of hallucination-detection papers."
   ],
   "mission": "To develop algorithms that let computers understand and produce human language, and to share the resulting tools with the research community.",
   "notice": null,
   "known_for": [
    "Nematus neural machine translation toolkit",
    "XSum extreme summarization dataset",
    "Leading academic NLP faculty"
   ],
   "focus": [
    "NLP",
    "machine translation",
    "summarization"
   ],
   "contributions": [
    {
     "title": "Nematus repository created",
     "note": "The open-source neural machine translation system was first published on GitHub.",
     "year": "2016"
    },
    {
     "title": "Nematus toolkit paper",
     "note": "Sennrich, Firat, Cho, Birch, Haddow and colleagues described the toolkit on arXiv.",
     "year": "2017"
    },
    {
     "title": "XSum dataset",
     "note": "Narayan, Cohen and Lapata introduced topic-aware convolutional networks and the extreme summarization dataset.",
     "year": "2018"
    },
    {
     "title": "Awesome hallucination detection list",
     "note": "The group opened a curated list of papers on hallucination detection that became its most starred repository.",
     "year": "2023"
    },
    {
     "title": "MMLongBench",
     "note": "A benchmark for long-context vision-language models was released by the group.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "awesome-hallucination-detection",
     "url": "https://github.com/EdinburghNLP/awesome-hallucination-detection",
     "note": "List of papers on hallucination detection in LLMs. (1,141 stars)"
    },
    {
     "name": "nematus",
     "url": "https://github.com/EdinburghNLP/nematus",
     "note": "Open-Source Neural Machine Translation in Tensorflow (805 stars)"
    },
    {
     "name": "XSum",
     "url": "https://github.com/EdinburghNLP/XSum",
     "note": "Topic-Aware Convolutional Neural Networks for Extreme Summarization (378 stars)"
    },
    {
     "name": "Refresh",
     "url": "https://github.com/EdinburghNLP/Refresh",
     "note": "Ranking Sentences for Extractive Summarization with Reinforcement Learning (275 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "EdinburghNLP/nunavut-hansard-eng",
     "url": "https://huggingface.co/EdinburghNLP/nunavut-hansard-eng",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "EdinburghNLP/nunavut-hansard-iku",
     "url": "https://huggingface.co/EdinburghNLP/nunavut-hansard-iku",
     "kind": "model",
     "note": "Model with 0 likes on the Hub."
    },
    {
     "name": "EdinburghNLP/xsum",
     "url": "https://huggingface.co/datasets/EdinburghNLP/xsum",
     "kind": "dataset",
     "note": "Dataset with 157 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Nematus: a Toolkit for Neural Machine Translation",
     "url": "https://arxiv.org/abs/1703.04357",
     "year": "2017"
    },
    {
     "title": "Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization",
     "url": "https://arxiv.org/abs/1808.08745",
     "year": "2018"
    }
   ],
   "notable_releases": [
    {
     "name": "Nematus",
     "url": "https://github.com/EdinburghNLP/nematus",
     "note": "Open-source neural machine translation toolkit"
    },
    {
     "name": "XSum dataset",
     "url": "https://huggingface.co/datasets/EdinburghNLP/xsum",
     "note": "Extreme summarisation dataset, with 150+ likes on the Hub"
    },
    {
     "name": "awesome-hallucination-detection",
     "url": "https://github.com/EdinburghNLP/awesome-hallucination-detection",
     "note": "Curated list of LLM hallucination detection papers"
    },
    {
     "name": "Refresh",
     "url": "https://github.com/EdinburghNLP/Refresh",
     "note": "Extractive summarisation with reinforcement learning"
    }
   ],
   "openness": "Code and datasets are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Mirella Lapata",
     "role": "Professor of natural language processing",
     "url": "https://en.wikipedia.org/wiki/Mirella_Lapata",
     "note": null
    },
    {
     "name": "Pasquale Minervini",
     "role": "Faculty member",
     "url": "https://edinburghnlp.inf.ed.ac.uk/index.php/dt_team/pasquale-minervini/",
     "note": null
    },
    {
     "name": "Edoardo Ponti",
     "role": "Faculty member",
     "url": "https://edinburghnlp.inf.ed.ac.uk/index.php/dt_team/edoardo-ponti/",
     "note": null
    },
    {
     "name": "Beatrice Alex",
     "role": "Faculty member",
     "url": "https://edinburghnlp.inf.ed.ac.uk/index.php/dt_team/alex-beatrice/",
     "note": null
    },
    {
     "name": "Sharon Goldwater",
     "role": "Faculty member",
     "url": "https://huggingface.co/sgwater",
     "note": null
    },
    {
     "name": "Jeff Dalton",
     "role": "Faculty member",
     "url": "https://edinburghnlp.inf.ed.ac.uk/index.php/dt_team/jeff-dalton/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://edinburghnlp.inf.ed.ac.uk",
    "https://huggingface.co/EdinburghNLP",
    "https://github.com/EdinburghNLP",
    "https://arxiv.org/abs/1703.04357"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "edinburghnlp",
   "logo": true
  },
  {
   "orig_id": "eleutherai",
   "name": "EleutherAI",
   "alias": null,
   "full_name": "EleutherAI",
   "type": "nonprofit",
   "founded": 2020,
   "country": "United States",
   "city": "Alexandria",
   "lat": 38.8048,
   "lon": -77.0469,
   "website": "https://www.eleuther.ai",
   "huggingface": "https://huggingface.co/EleutherAI",
   "github": "https://github.com/EleutherAI",
   "x": "https://x.com/AIEleuther",
   "discord": "https://discord.gg/eleutherai",
   "linkedin": null,
   "contact_email": "contact@eleuther.ai",
   "summary": "EleutherAI is a non-profit research institute that trains and releases open models, datasets and tools. It began in 2020 as a Discord server and incorporated in 2023.",
   "tagline": "EleutherAI is a non-profit research institute that grew from a Discord server into a leading source of open language models, datasets and evaluation tools.",
   "about": [
    "EleutherAI began in July 2020 as a Discord server under the working name LibreAI. Connor Leahy, Leo Gao and Sid Black are named as founders, and the group was renamed EleutherAI later that month. In early 2023 it was incorporated as a non-profit research institute.",
    "It first used compute from Google's TPU Research Cloud and later from CoreWeave. The group released The Pile in December 2020, followed by the GPT-Neo, GPT-J and GPT-NeoX-20B models, and the Pythia suite for studying training dynamics.",
    "Today it reports about fifteen full- and part-time staff along with regular volunteers. Its lm-evaluation-harness, the most starred repository in its GitHub organization, is used widely to benchmark language models."
   ],
   "mission": "To advance interpretability and alignment research on open foundation models, keep such models open to study beyond a few companies, and educate the public about their capabilities and risks.",
   "notice": null,
   "known_for": [
    "The Pile dataset",
    "GPT-Neo, GPT-J and GPT-NeoX-20B",
    "lm-evaluation-harness and Pythia"
   ],
   "focus": [
    "language models",
    "evaluation",
    "interpretability",
    "datasets",
    "distributed"
   ],
   "contributions": [
    {
     "title": "The Pile released",
     "note": "An 800GB dataset of diverse text for language modeling, authored by Gao, Biderman, Black and others.",
     "year": "2020"
    },
    {
     "title": "GPT-Neo",
     "note": "The first GPT-Neo models of 125M, 1.3B and 2.7B parameters were published in March 2021.",
     "year": "2021"
    },
    {
     "title": "GPT-J-6B",
     "note": "A six-billion-parameter open model was released in June 2021.",
     "year": "2021"
    },
    {
     "title": "GPT-NeoX-20B",
     "note": "A 20-billion-parameter model trained with CoreWeave compute was released in February 2022.",
     "year": "2022"
    },
    {
     "title": "Pythia suite and incorporation",
     "note": "Pythia provided checkpoints for analyzing models across training and scaling, and the group incorporated as a non-profit institute.",
     "year": "2023"
    },
    {
     "title": "Common Pile v0.1",
     "note": "A training dataset built to exclude copyrighted material was released in June 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "lm-evaluation-harness",
     "url": "https://github.com/EleutherAI/lm-evaluation-harness",
     "note": "Most-starred repo; the standard few-shot LLM evaluation framework."
    },
    {
     "name": "gpt-neo",
     "url": "https://github.com/EleutherAI/gpt-neo",
     "note": "Model-parallel GPT-2/GPT-3-style training library from the early GPT-Neo era."
    },
    {
     "name": "gpt-neox",
     "url": "https://github.com/EleutherAI/gpt-neox",
     "note": "Large-scale autoregressive training library behind GPT-NeoX-20B and Pythia."
    },
    {
     "name": "pythia",
     "url": "https://github.com/EleutherAI/pythia",
     "note": "Hub for the Pythia suite and learning-dynamics research."
    }
   ],
   "hf_highlights": [
    {
     "name": "pythia-160m",
     "url": "https://huggingface.co/EleutherAI/pythia-160m",
     "kind": "model",
     "note": "Small Pythia model, the org's most downloaded."
    },
    {
     "name": "gpt-neo-125m",
     "url": "https://huggingface.co/EleutherAI/gpt-neo-125m",
     "kind": "model",
     "note": "GPT-Neo 125M open model."
    },
    {
     "name": "hendrycks_math",
     "url": "https://huggingface.co/datasets/EleutherAI/hendrycks_math",
     "kind": "dataset",
     "note": "MATH benchmark packaged for the evaluation harness."
    },
    {
     "name": "lambada_openai",
     "url": "https://huggingface.co/datasets/EleutherAI/lambada_openai",
     "kind": "dataset",
     "note": "LAMBADA evaluation set used by the harness."
    }
   ],
   "papers": [
    {
     "title": "The Pile: An 800GB Dataset of Diverse Text for Language Modeling",
     "url": "https://arxiv.org/abs/2101.00027",
     "year": "2020"
    },
    {
     "title": "Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling",
     "url": "https://arxiv.org/abs/2304.01373",
     "year": "2023"
    },
    {
     "title": "GPT-NeoX-20B: An Open-Source Autoregressive Language Model",
     "url": "https://arxiv.org/abs/2204.06745",
     "year": "2022"
    },
    {
     "title": "Lessons from the Trenches on Reproducible Evaluation of Language Models",
     "url": "https://arxiv.org/abs/2405.14782",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "lm-evaluation-harness",
     "url": "https://github.com/EleutherAI/lm-evaluation-harness",
     "note": "Evaluation framework"
    },
    {
     "name": "GPT-NeoX",
     "url": "https://github.com/EleutherAI/gpt-neox",
     "note": "Training library"
    },
    {
     "name": "Pythia",
     "url": "https://huggingface.co/EleutherAI/pythia-160m",
     "note": "Model suite"
    }
   ],
   "openness": "Open weights, datasets and code.",
   "key_people": [
    {
     "name": "Connor Leahy",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/EleutherAI",
     "note": null
    },
    {
     "name": "Leo Gao",
     "role": "Co-founder, lead author of The Pile",
     "url": "https://en.wikipedia.org/wiki/EleutherAI",
     "note": null
    },
    {
     "name": "Sid Black",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/EleutherAI",
     "note": null
    },
    {
     "name": "Stella Biderman",
     "role": "Leader of the incorporated institute, Pythia lead author",
     "url": "https://github.com/StellaAthena",
     "note": null
    },
    {
     "name": "Quentin Anthony",
     "role": "Researcher, Pythia co-author",
     "url": "https://github.com/Quentin-Anthony",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.eleuther.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.eleuther.ai/about",
    "https://huggingface.co/EleutherAI",
    "https://github.com/EleutherAI"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "eleutherai",
   "logo": true
  },
  {
   "orig_id": "ellamind",
   "name": "ellamind",
   "alias": null,
   "full_name": "ellamind",
   "type": "startup",
   "founded": null,
   "country": "Germany",
   "city": "",
   "lat": 51.1657,
   "lon": 10.4515,
   "website": "https://ellamind.com",
   "huggingface": "https://huggingface.co/ellamind",
   "github": "https://github.com/ellamind",
   "x": "https://x.com/ellamindAI",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/ellamind",
   "contact_email": null,
   "summary": "German company working on AI evaluation, data and agents. It open-sources tooling for evaluation and inference and publishes models and datasets.",
   "tagline": "A German AI company, based in Bremen and Berlin, that builds evaluation tools and open-source infrastructure for language models.",
   "about": [
    "ellamind GmbH works on evaluation, data and environments for AI systems and helps regulated industries such as health insurance, banking and public administration bring them into production. Its products include elluminate, an agentic evaluation platform, and an accessibility testing agent.",
    "The company open-sources base-eval, curated lm-evaluation-harness task configurations for English and German base models, and inference-hive, distributed LLM inference for SLURM clusters, both under Apache-2.0. It also publishes propella-1, a family of small multilingual annotation models, and sui-1, a 24B-parameter summarization model with inline citations, on Hugging Face."
   ],
   "mission": "ellamind focuses on making AI systems measurable and dependable for organisations in regulated sectors, and shares part of its tooling as open source.",
   "notice": null,
   "known_for": [
    "Open evaluation tooling for German models",
    "inference-hive for SLURM clusters",
    "propella-1 annotation models"
   ],
   "focus": [
    "evaluation",
    "datasets",
    "tooling"
   ],
   "contributions": [],
   "github_highlights": [
    {
     "name": "autorag",
     "url": "https://github.com/ellamind/autorag",
     "note": "Automatic hyperparameter optimisation of RAG pipelines (8 stars)"
    },
    {
     "name": "base-eval",
     "url": "https://github.com/ellamind/base-eval",
     "note": "Base model evaluation (8 stars)"
    },
    {
     "name": "eval-hive",
     "url": "https://github.com/ellamind/eval-hive",
     "note": "Evaluation toolkit (7 stars)"
    },
    {
     "name": "inference-hive",
     "url": "https://github.com/ellamind/inference-hive",
     "note": "Toolkit for distributed LLM inference on SLURM clusters (4 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "ellamind/propella-1-4b",
     "url": "https://huggingface.co/ellamind/propella-1-4b",
     "kind": "model",
     "note": "4B document annotation model, ellamind's most downloaded model."
    },
    {
     "name": "ellamind/propella-1-0.6b",
     "url": "https://huggingface.co/ellamind/propella-1-0.6b",
     "kind": "model",
     "note": "0.6B document annotation model."
    },
    {
     "name": "ellamind/arc-multilingual",
     "url": "https://huggingface.co/datasets/ellamind/arc-multilingual",
     "kind": "dataset",
     "note": "Multilingual ARC benchmark."
    },
    {
     "name": "ellamind/base-eval",
     "url": "https://huggingface.co/spaces/ellamind/base-eval",
     "kind": "space",
     "note": "Base model evaluation space."
    }
   ],
   "papers": [
    {
     "title": "propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale",
     "url": "https://arxiv.org/abs/2602.12414",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "propella-1-4b",
     "url": "https://huggingface.co/ellamind/propella-1-4b",
     "note": "Document annotation model"
    },
    {
     "name": "base-eval",
     "url": "https://github.com/ellamind/base-eval",
     "note": "Evaluation tooling"
    },
    {
     "name": "inference-hive",
     "url": "https://github.com/ellamind/inference-hive",
     "note": "Toolkit for distributed LLM inference on SLURM"
    }
   ],
   "openness": "Code is mostly MIT or Apache-2.0; models and datasets are public on Hugging Face.",
   "key_people": [],
   "discord_note": null,
   "contact_note": "The company site has a contact page and a demo booking form.",
   "sources": [
    "https://ellamind.com",
    "https://huggingface.co/ellamind",
    "https://github.com/ellamind"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "ellamind",
   "mlat": 51.13034,
   "mlon": 10.4515,
   "logo": true
  },
  {
   "orig_id": "ellis-alicante",
   "name": "ELLIS Alicante",
   "alias": null,
   "full_name": "ELLIS Alicante",
   "type": "nonprofit",
   "founded": 2020,
   "country": "Spain",
   "city": "Alicante",
   "lat": 38.3452,
   "lon": -0.481,
   "website": "https://ellisalicante.org",
   "huggingface": null,
   "github": "https://github.com/ellisalicante",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "ELLIS Alicante is a non-profit research unit of the ELLIS network focused on human-centric AI, with research code and benchmarks on GitHub.",
   "tagline": "ELLIS Alicante is a non-profit unit of the European ELLIS network focused on ethical, human-centric AI, directed by Nuria Oliver.",
   "about": [
    "ELLIS Alicante was founded in 2020 as a foundation hosting a research unit of the ELLIS network in Spain. It resulted from a 2019 bid, led by Nuria Oliver, for Alicante to host an ELLIS unit, and she serves as its director.",
    "Its research addresses modeling human behavior at individual and societal scales, systems that interact naturally with people, and ethical AI covering fairness, privacy and transparency. It is funded by the Generalitat Valenciana, the European Regional Development Fund and partners including Intel and Fundación Banco Sabadell, and reports over 38 million euros in EU project budgets.",
    "Paper code and evaluation repositories are published on GitHub, including tutorials on graph rewiring and fairness, an open framework for social media AR filters, and evaluations of how language models handle mental-health crises."
   ],
   "mission": "To advance ethical, responsible and human-centric artificial intelligence that serves society.",
   "notice": null,
   "known_for": [
    "Human-centric and ethical AI research",
    "Directed by Nuria Oliver",
    "Open benchmarks on fairness and LLM safety"
   ],
   "focus": [
    "responsible AI"
   ],
   "contributions": [
    {
     "title": "Alicante wins an ELLIS unit",
     "note": "A bid led by Nuria Oliver brought a unit of the ELLIS network to Alicante.",
     "year": "2019"
    },
    {
     "title": "ELLIS Alicante foundation established",
     "note": "The foundation began operating and opened its GitHub organization in April 2020.",
     "year": "2020"
    },
    {
     "title": "OpenFilter framework",
     "note": "The group released a framework to democratize research access to social media AR filters, with Piera Riccio, Nuria Oliver and coauthors.",
     "year": "2022"
    },
    {
     "title": "Graph rewiring tutorial",
     "note": "A tutorial on graph rewiring and fairness was presented at the Learning on Graphs conference.",
     "year": "2022"
    },
    {
     "title": "Multilingual and multimodal red teaming",
     "note": "Code accompanying a paper on red teaming multimodal models was presented at the ACL 2025 LLM security workshop.",
     "year": "2025"
    },
    {
     "title": "LLM mental-health crisis evaluation",
     "note": "An evaluation of how large language models handle mental-health crises was published with open code.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "GraphRewiring-Tutorial",
     "url": "https://github.com/ellisalicante/GraphRewiring-Tutorial",
     "note": "This repository holds code and other relevant files for the Learning on Graphs 2022 (LoG) tutorial \"Graph Rewi (54 stars)"
    },
    {
     "name": "OpenFilter",
     "url": "https://github.com/ellisalicante/OpenFilter",
     "note": "This repository refers to the paper currently under review for the 36th Conference on Neural Information Proce (13 stars)"
    },
    {
     "name": "LLMs-Mental-Health-Crisis",
     "url": "https://github.com/ellisalicante/LLMs-Mental-Health-Crisis",
     "note": "JMIR MH 2026 - \"Between Help and Harm: An Evaluation of Mental Health Crisis Handling by LLMs\" Adrian Arnaiz-R (9 stars)"
    },
    {
     "name": "multilingual-multimodal-red-teaming",
     "url": "https://github.com/ellisalicante/multilingual-multimodal-red-teaming",
     "note": "Repository accompanying our paper titled \"Beyond Words: Multilingual and Multimodal Red Teaming of MLLMs\", pre (3 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "GraphRewiring-Tutorial",
     "url": "https://github.com/ellisalicante/GraphRewiring-Tutorial",
     "note": "Tutorial code from the Learning on Graphs conference"
    },
    {
     "name": "OpenFilter",
     "url": "https://github.com/ellisalicante/OpenFilter",
     "note": "Code for the OpenFilter paper"
    },
    {
     "name": "LLMs-Mental-Health-Crisis",
     "url": "https://github.com/ellisalicante/LLMs-Mental-Health-Crisis",
     "note": "Evaluation of mental health crisis handling by LLMs"
    },
    {
     "name": "Multilingual red teaming",
     "url": "https://github.com/ellisalicante/multilingual-multimodal-red-teaming",
     "note": "Code for multilingual and multimodal red teaming"
    }
   ],
   "openness": "Paper code and evaluation repositories are public on GitHub.",
   "key_people": [
    {
     "name": "Nuria Oliver",
     "role": "Director",
     "url": "https://en.wikipedia.org/wiki/Nuria_Oliver",
     "note": null
    },
    {
     "name": "Adrian Arnaiz-Rodriguez",
     "role": "Researcher",
     "url": "https://github.com/AdrianArnaiz",
     "note": null
    },
    {
     "name": "Piera Riccio",
     "role": "Researcher, OpenFilter first author",
     "url": "https://github.com/ellisalicante/OpenFilter",
     "note": "Profile not found; repository lists her as author"
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://ellisalicante.org",
    "https://github.com/ellisalicante"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "ellis-alicante",
   "logo": true
  },
  {
   "orig_id": "epfl-llm",
   "name": "EPFL LLM Team",
   "alias": "Meditron",
   "full_name": "EPFL LLM Team (Meditron)",
   "type": "academic",
   "founded": null,
   "country": "Switzerland",
   "city": "Lausanne",
   "lat": 46.5191,
   "lon": 6.5668,
   "website": "https://huggingface.co/epfl-llm",
   "huggingface": "https://huggingface.co/epfl-llm",
   "github": "https://github.com/epfLLM",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "EPFL team behind Meditron, open medical LLMs, and the Megatron-LLM trainer.",
   "tagline": "The EPFL LLM team built Meditron, an open suite of medical language models, and the Megatron-LLM trainer used to produce them.",
   "about": [
    "The EPFL LLM team is a group at the Swiss federal institute in Lausanne that publishes under the epfLLM GitHub and epfl-llm Hugging Face organizations. Its Megatron-LLM repository, created in July 2023, adapts NVIDIA's Megatron-LM trainer for distributed training of open models.",
    "In November 2023 the team released MEDITRON, 7B and 70B-parameter medical models built on Llama-2 and further pretrained on PubMed papers, abstracts and international medical guidelines. The paper reports that MEDITRON-70B outperforms GPT-3.5 and Med-PaLM on four benchmarks, and the team released both model weights and the corpus-building code."
   ],
   "mission": "To provide openly available medical language models and training tools so that clinical AI can be studied and improved in the open.",
   "notice": null,
   "known_for": [
    "Meditron open medical LLMs",
    "Megatron-LLM distributed trainer",
    "Open weights and corpus-building code"
   ],
   "focus": [
    "medical LLMs",
    "training infrastructure"
   ],
   "contributions": [
    {
     "title": "GitHub organization created",
     "note": "The epfLLM organization was opened to host the team's open training code.",
     "year": "2023"
    },
    {
     "title": "Megatron-LLM",
     "note": "A distributed trainer for large language models was published on GitHub.",
     "year": "2023"
    },
    {
     "title": "MEDITRON-70B paper",
     "note": "Chen, Bosselut, Jaggi and colleagues released 7B and 70B medical models and the code to rebuild their pretraining corpus.",
     "year": "2023"
    },
    {
     "title": "Meditron project website",
     "note": "The team launched the official Meditron project site, with the repository gaining over 2,000 stars.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "meditron",
     "url": "https://github.com/epfLLM/meditron",
     "note": "Meditron is a suite of open-source medical Large Language Models (LLMs). (2,221 stars)"
    },
    {
     "name": "Megatron-LLM",
     "url": "https://github.com/epfLLM/Megatron-LLM",
     "note": "distributed trainer for LLMs (591 stars)"
    },
    {
     "name": "meditron-epfl",
     "url": "https://github.com/epfLLM/meditron-epfl",
     "note": "Official project website for Meditron, developed by the EPFL LLM team. (2 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "epfl-llm/meditron-7b",
     "url": "https://huggingface.co/epfl-llm/meditron-7b",
     "kind": "model",
     "note": "Model with 331 likes on the Hub."
    },
    {
     "name": "epfl-llm/meditron-70b",
     "url": "https://huggingface.co/epfl-llm/meditron-70b",
     "kind": "model",
     "note": "Model with 274 likes on the Hub."
    },
    {
     "name": "epfl-llm/guidelines",
     "url": "https://huggingface.co/datasets/epfl-llm/guidelines",
     "kind": "dataset",
     "note": "Dataset with 167 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "MEDITRON-70B: Scaling Medical Pretraining for Large Language Models",
     "url": "https://arxiv.org/abs/2311.16079",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "meditron-7b",
     "url": "https://huggingface.co/epfl-llm/meditron-7b",
     "note": "Open medical LLM"
    },
    {
     "name": "meditron-70b",
     "url": "https://huggingface.co/epfl-llm/meditron-70b",
     "note": "70B medical LLM"
    },
    {
     "name": "guidelines",
     "url": "https://huggingface.co/datasets/epfl-llm/guidelines",
     "note": "Clinical guidelines corpus"
    },
    {
     "name": "Megatron-LLM",
     "url": "https://github.com/epfLLM/Megatron-LLM",
     "note": "Distributed LLM trainer"
    }
   ],
   "openness": "Weights, data and training code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Antoine Bosselut",
     "role": "Senior author of MEDITRON",
     "url": "https://huggingface.co/atcbosselut",
     "note": null
    },
    {
     "name": "Martin Jaggi",
     "role": "Senior author of MEDITRON",
     "url": "https://huggingface.co/mjaggi",
     "note": null
    },
    {
     "name": "Mary-Anne Hartley",
     "role": "Senior author of MEDITRON",
     "url": "https://huggingface.co/Yale-LiGHT",
     "note": null
    },
    {
     "name": "Zeming Chen",
     "role": "Lead author of MEDITRON",
     "url": "https://huggingface.co/zechen-nlp",
     "note": null
    },
    {
     "name": "Alexandre Sallinen",
     "role": "MEDITRON co-author",
     "url": "https://huggingface.co/alexsallinen",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the project website.",
   "sources": [
    "https://huggingface.co/epfl-llm",
    "https://huggingface.co/epfl-llm",
    "https://github.com/epfLLM",
    "https://arxiv.org/abs/2311.16079"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "epfl-llm-team",
   "mlat": 46.48374,
   "mlon": 6.5668,
   "logo": true
  },
  {
   "orig_id": "epoch-ai",
   "name": "Epoch AI",
   "alias": null,
   "full_name": "Epoch AI",
   "type": "nonprofit",
   "founded": 2022,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://epoch.ai",
   "huggingface": null,
   "github": "https://github.com/epoch-research",
   "x": "https://x.com/EpochAIResearch",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Research organisation that tracks AI compute, data, algorithms and hardware trends, publishing public datasets and analysis code.",
   "tagline": "A research organization that publishes open datasets and analysis on AI compute, hardware, investment and data centers.",
   "about": [
    "Epoch AI began in 2021 as a group of volunteers and became a formal organization after its 2022 paper on training compute trends. It describes its aim as helping people understand AI from a neutral perspective grounded in the best available evidence.",
    "Its public databases cover ML hardware, training compute estimates for AI models, AI investment and companies, GPU clusters and AI data centers. It also publishes analysis, chip sales data and polling on AI usage."
   ],
   "mission": "Epoch AI seeks to improve society's understanding of the drivers, progress and impact of AI, using data to inform decisions about the technology.",
   "notice": null,
   "known_for": [
    "Training compute estimates for AI models",
    "Public databases on GPU clusters and data centers",
    "Evidence-based AI trend analysis"
   ],
   "focus": [
    "AI trends",
    "datasets",
    "forecasting",
    "benchmarking"
   ],
   "contributions": [
    {
     "title": "Volunteer group forms",
     "note": "Epoch AI began as volunteers collecting and analyzing data on AI progress.",
     "year": "2021"
    },
    {
     "title": "Training compute trends paper",
     "note": "The 2022 paper on compute trends led to the founding of the organization.",
     "year": "2022"
    },
    {
     "title": "Public databases",
     "note": "Maintains open databases on ML hardware, model training compute, AI investment, GPU clusters and data centers.",
     "year": "Ongoing"
    },
    {
     "title": "Benchmarks and evaluations",
     "note": "Runs independent evaluations and builds benchmarks, including FrontierMath.",
     "year": "Ongoing"
    }
   ],
   "github_highlights": [
    {
     "name": "MirrorCode",
     "url": "https://github.com/epoch-research/MirrorCode",
     "note": "Public repository for MirrorCode"
    },
    {
     "name": "Compute-Trends",
     "url": "https://github.com/epoch-research/Compute-Trends",
     "note": "Supplementary material for compute-trends paper"
    },
    {
     "name": "training-cost-trends",
     "url": "https://github.com/epoch-research/training-cost-trends",
     "note": "Training cost trends analysis"
    },
    {
     "name": "data-stock",
     "url": "https://github.com/epoch-research/data-stock",
     "note": "Models for data stocks and dataset sizes"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Compute Trends Across Three Eras of Machine Learning",
     "url": "https://arxiv.org/abs/2202.05924",
     "year": "2022"
    },
    {
     "title": "Will we run out of data? Limits of LLM scaling based on human-generated data",
     "url": "https://arxiv.org/abs/2211.04325",
     "year": "2022"
    },
    {
     "title": "Algorithmic progress in language models",
     "url": "https://arxiv.org/abs/2403.05812",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "MirrorCode",
     "url": "https://github.com/epoch-research/MirrorCode",
     "note": "Public repository for MirrorCode"
    },
    {
     "name": "Compute-Trends",
     "url": "https://github.com/epoch-research/Compute-Trends",
     "note": "Supplementary material for compute-trends paper"
    },
    {
     "name": "training-cost-trends",
     "url": "https://github.com/epoch-research/training-cost-trends",
     "note": "Training cost trends analysis"
    }
   ],
   "openness": "Datasets published openly on epoch.ai; analysis code on GitHub, mostly MIT.",
   "key_people": [
    {
     "name": "Jaime Sevilla",
     "role": "CEO and board member",
     "url": "https://epoch.ai/about/team/jaime-sevilla",
     "note": null
    },
    {
     "name": "María de la Lama",
     "role": "COO and board member",
     "url": "https://epoch.ai/about/team/maria-de-la-lama",
     "note": null
    },
    {
     "name": "Yafah Edelman",
     "role": "Chief Scientific Officer",
     "url": "https://epoch.ai/about/team/yafah-edelman",
     "note": null
    },
    {
     "name": "Bret Sepulveda",
     "role": "Chief Technology Officer",
     "url": "https://epoch.ai/about/team/bret-sepulveda",
     "note": null
    },
    {
     "name": "Greg Burnham",
     "role": "Head of Benchmarks",
     "url": "https://epoch.ai/about/team/greg-burnham",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website; no organisation email printed on the pages opened.",
   "sources": [
    "https://epoch.ai",
    "https://epoch.ai/about",
    "https://github.com/epoch-research"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "epoch-ai",
   "mlat": 37.81932,
   "mlon": -122.27837,
   "logo": true
  },
  {
   "orig_id": "essential-ai",
   "name": "Essential AI",
   "alias": null,
   "full_name": "Essential AI",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.essential.ai",
   "huggingface": "https://huggingface.co/EssentialAI",
   "github": "https://github.com/Essential-AI",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@essential.ai",
   "summary": "San Francisco lab releasing open models (Rnj-1) and the Essential-Web dataset.",
   "tagline": "A San Francisco lab led by Ashish Vaswani that builds an open platform for AI and has released the Rnj-1 model and the Essential-Web dataset.",
   "about": [
    "Essential AI is led by CEO Ashish Vaswani and describes itself as building an open platform to fuel and accelerate AI breakthroughs globally. Its research output covers pretraining data, optimizers and model releases.",
    "The lab published Essential-Web, a 24-trillion-token organized web dataset, in June 2025 and a series of papers and posts on the Muon optimizer in May and June 2025. In December 2025 it announced Rnj-1, which it calls its first instrument of intelligence."
   ],
   "mission": "Essential AI aims to build an open platform that accelerates AI progress, and it shares datasets and research as part of that effort.",
   "notice": null,
   "known_for": [
    "Essential-Web 24T-token dataset",
    "Rnj-1 open model",
    "Research on the Muon optimizer"
   ],
   "focus": [
    "language models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Practical efficiency of Muon",
     "note": "Research paper on the Muon optimizer for pretraining.",
     "year": "May 2025"
    },
    {
     "title": "Essential-Web v1.0",
     "note": "Release of a 24-trillion-token organized web dataset.",
     "year": "June 2025"
    },
    {
     "title": "Rnj-1",
     "note": "Announcement of base and instruction-tuned models.",
     "year": "December 2025"
    }
   ],
   "github_highlights": [
    {
     "name": "eai-taxonomy",
     "url": "https://github.com/Essential-AI/eai-taxonomy",
     "note": "Taxonomy for web documents."
    },
    {
     "name": "reflection",
     "url": "https://github.com/Essential-AI/reflection",
     "note": "Open repo from Essential AI."
    },
    {
     "name": "rnj-1-assets",
     "url": "https://github.com/Essential-AI/rnj-1-assets",
     "note": "Assets for Rnj-1."
    }
   ],
   "hf_highlights": [
    {
     "name": "eai-distill-0.5b",
     "url": "https://huggingface.co/EssentialAI/eai-distill-0.5b",
     "kind": "model",
     "note": "Document classifier distilled for labelling."
    },
    {
     "name": "rnj-1-instruct",
     "url": "https://huggingface.co/EssentialAI/rnj-1-instruct",
     "kind": "model",
     "note": "Rnj-1 instruct model."
    },
    {
     "name": "essential-web-v1.0",
     "url": "https://huggingface.co/datasets/EssentialAI/essential-web-v1.0",
     "kind": "dataset",
     "note": "Labelled web dataset."
    }
   ],
   "papers": [
    {
     "title": "Essential-Web v1.0: 24T tokens of organized web data",
     "url": "https://arxiv.org/abs/2506.14111",
     "year": "2025"
    },
    {
     "title": "Practical Efficiency of Muon for Pretraining",
     "url": "https://arxiv.org/abs/2505.02222",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "rnj-1-instruct",
     "url": "https://huggingface.co/EssentialAI/rnj-1-instruct",
     "note": "Open model"
    },
    {
     "name": "essential-web-v1.0",
     "url": "https://huggingface.co/datasets/EssentialAI/essential-web-v1.0",
     "note": "Taxonomy-labelled web data"
    }
   ],
   "openness": "Open weights and data.",
   "key_people": [
    {
     "name": "Ashish Vaswani",
     "role": "CEO",
     "url": "https://www.essential.ai",
     "note": "Author of the Rnj-1 announcement; co-author of the Transformer paper."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.essential.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.essential.ai",
    "https://huggingface.co/EssentialAI"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "essential-ai",
   "mlat": 37.84224,
   "mlon": -122.29374,
   "logo": true
  },
  {
   "orig_id": "eurollm",
   "name": "EuroLLM",
   "alias": "UTTER",
   "full_name": "EuroLLM (UTTER)",
   "type": "collective",
   "founded": null,
   "country": "Portugal",
   "city": "Lisbon",
   "lat": 38.7223,
   "lon": -9.1393,
   "website": "https://eurollm.io",
   "huggingface": "https://huggingface.co/utter-project",
   "github": "https://github.com/utter-project",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "EuroLLM is a series of open multilingual language models for European languages, developed by a consortium of European research institutions and companies under the UTTER project.",
   "tagline": "A series of open multilingual language models covering the 24 official EU languages, built by a European research and industry consortium.",
   "about": [
    "EuroLLM is developed under the Horizon Europe UTTER project by a consortium that includes Instituto Superior Tecnico, the University of Edinburgh, Instituto de Telecomunicacoes, Universite Paris-Saclay, Unbabel, Sorbonne University, Naver Labs and the University of Amsterdam. The project is funded by Horizon Europe, the European Research Council and EuroHPC.",
    "The models support all 24 official EU languages and training data covers 35 languages. Released models include EuroLLM-1.7B for edge devices, EuroLLM-9B and EuroLLM-22B with a 32K context, in base and instruct versions, available on Hugging Face."
   ],
   "mission": "EuroLLM aims to provide fully open language models that serve Europe's linguistic diversity, rather than relying on models centered on English.",
   "notice": null,
   "known_for": [
    "Open models for all 24 EU languages",
    "EuroLLM-9B and 22B releases",
    "UTTER Horizon Europe project"
   ],
   "focus": [
    "language models",
    "multilingual",
    "European languages",
    "translation"
   ],
   "contributions": [
    {
     "title": "EuroHPC compute grant",
     "note": "The project used a EuroHPC resource grant (EHPC-EXT-2023E01-042) for training.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "mHuBERT-147-scripts",
     "url": "https://github.com/utter-project/mHuBERT-147-scripts",
     "note": "Collection of scripts from  mHuBERT-147. (35 stars)"
    },
    {
     "name": "fairseq",
     "url": "https://github.com/utter-project/fairseq",
     "note": "This is a fork of the original fairseq repository (version 0.12.2) with added classes for training mHuBERT-147 (21 stars)"
    },
    {
     "name": "SpireLM",
     "url": "https://github.com/utter-project/SpireLM",
     "note": "Open-source repository (10 stars)"
    },
    {
     "name": "ELITR-Bench",
     "url": "https://github.com/utter-project/ELITR-Bench",
     "note": "Benchmark for the evaluation of long-context LLMs on meeting transcripts (6 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "utter-project/mHuBERT-147",
     "url": "https://huggingface.co/utter-project/mHuBERT-147",
     "kind": "model",
     "note": "Model: 33,893 downloads, 108 likes."
    },
    {
     "name": "utter-project/EuroLLM-1.7B-Instruct",
     "url": "https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct",
     "kind": "model",
     "note": "Model: 19,733 downloads, 104 likes."
    },
    {
     "name": "utter-project/EuroWeb-2512",
     "url": "https://huggingface.co/datasets/utter-project/EuroWeb-2512",
     "kind": "dataset",
     "note": "Dataset: 19,106 downloads, 6 likes."
    },
    {
     "name": "utter-project/EuroLLM-1.7B",
     "url": "https://huggingface.co/utter-project/EuroLLM-1.7B",
     "kind": "model",
     "note": "Model: 12,462 downloads, 123 likes."
    }
   ],
   "papers": [
    {
     "title": "EuroLLM-9B: Technical Report",
     "url": "https://arxiv.org/abs/2506.04079",
     "year": "2025"
    },
    {
     "title": "EuroLLM: Multilingual Language Models for Europe",
     "url": "https://arxiv.org/abs/2409.16235",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "EuroLLM-9B-Instruct",
     "url": "https://huggingface.co/utter-project/EuroLLM-9B-Instruct",
     "note": "Instruction-tuned 9B multilingual model"
    },
    {
     "name": "EuroLLM-9B",
     "url": "https://huggingface.co/utter-project/EuroLLM-9B",
     "note": "Base 9B model"
    },
    {
     "name": "EuroBlocks SFT",
     "url": "https://huggingface.co/datasets/utter-project/EuroBlocks-SFT-2512",
     "note": "Instruction-tuning data"
    }
   ],
   "openness": "Model weights and instruction data are public on Hugging Face; the technical report is open.",
   "key_people": [
    {
     "name": "Andre Martins",
     "role": "Project lead, Instituto Superior Tecnico",
     "url": "https://eurollm.io",
     "note": "Listed among key people on the project site."
    },
    {
     "name": "Alexandra Birch",
     "role": "University of Edinburgh",
     "url": "https://eurollm.io",
     "note": "Listed among key people on the project site."
    },
    {
     "name": "Pierre Colombo",
     "role": "Key person named on the project site",
     "url": "https://eurollm.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the project website or Hugging Face org.",
   "sources": [
    "https://eurollm.io",
    "https://huggingface.co/utter-project",
    "https://github.com/utter-project",
    "https://arxiv.org/abs/2506.04079"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "eurollm",
   "mlat": 38.68694,
   "mlon": -9.1393,
   "logo": true
  },
  {
   "orig_id": "explosion",
   "name": "Explosion",
   "alias": null,
   "full_name": "Explosion",
   "type": "startup",
   "founded": 2016,
   "country": "Germany",
   "city": "Berlin",
   "lat": 52.52,
   "lon": 13.405,
   "website": "https://explosion.ai",
   "huggingface": "https://huggingface.co/explosion",
   "github": "https://github.com/explosion",
   "x": null,
   "discord": null,
   "linkedin": "https://linkedin.com/company/explosion-ai",
   "contact_email": "contact@explosion.ai",
   "summary": "Berlin software company that created spaCy and Thinc, open-source libraries for NLP and machine learning.",
   "tagline": "A Berlin software company, founded by Matthew Honnibal and Ines Montani, that created the spaCy open-source NLP library and the Prodigy annotation tool.",
   "about": [
    "Explosion is a Berlin company founded by Matthew Honnibal, its CTO, and Ines Montani, its CEO. It makes spaCy, an open-source library for natural language processing, and Prodigy, an annotation tool, alongside newer products such as Ellf, an assistant for agentic NLP development.",
    "The company maintains its software in the open and describes an employee pool with a stake in the business. Its Hugging Face and GitHub organisations host spaCy models, Thinc and related libraries."
   ],
   "mission": "Explosion builds practical developer tools for natural language processing and keeps the core libraries open source.",
   "notice": null,
   "known_for": [
    "spaCy NLP library",
    "Prodigy annotation tool",
    "Founder-led independent software company"
   ],
   "focus": [
    "NLP",
    "tooling",
    "machine learning libraries"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Explosion was founded in Berlin by Matthew Honnibal and Ines Montani.",
     "year": "2016"
    },
    {
     "title": "SignalFire investment",
     "note": "Sold 5% of the company to SignalFire for $6 million.",
     "year": "2021"
    },
    {
     "title": "Follow-on investment",
     "note": "SignalFire invested a further $3 million.",
     "year": "2022"
    },
    {
     "title": "Return to independence",
     "note": "Described its return to running as an independent, self-sufficient company with no further venture funding sought.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "spaCy",
     "url": "https://github.com/explosion/spaCy",
     "note": "💫 Industrial-strength Natural Language Processing (NLP) in Python (34,062 stars)"
    },
    {
     "name": "thinc",
     "url": "https://github.com/explosion/thinc",
     "note": "🔮 A refreshing functional take on deep learning, compatible with your favorite libraries (2,891 stars)"
    },
    {
     "name": "spacy-course",
     "url": "https://github.com/explosion/spacy-course",
     "note": "👩‍🏫 Advanced NLP with spaCy: A free online course (2,433 stars)"
    },
    {
     "name": "spacy-models",
     "url": "https://github.com/explosion/spacy-models",
     "note": "💫  Models for the spaCy Natural Language Processing (NLP) library (1,920 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "explosion/en_textcat_goemotions",
     "url": "https://huggingface.co/explosion/en_textcat_goemotions",
     "kind": "model",
     "note": "Model: 31 downloads, 11 likes."
    },
    {
     "name": "explosion/en_healthsea",
     "url": "https://huggingface.co/explosion/en_healthsea",
     "kind": "model",
     "note": "Model: 15 downloads, 5 likes."
    },
    {
     "name": "explosion/ner-fashion-brands",
     "url": "https://huggingface.co/datasets/explosion/ner-fashion-brands",
     "kind": "dataset",
     "note": "Dataset: 39 downloads, 1 likes."
    },
    {
     "name": "explosion/da_udv25_danishddt_trf",
     "url": "https://huggingface.co/explosion/da_udv25_danishddt_trf",
     "kind": "model",
     "note": "Model: 14 downloads, 0 likes."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "spaCy",
     "url": "https://github.com/explosion/spaCy",
     "note": "Industrial-strength NLP library"
    },
    {
     "name": "Thinc",
     "url": "https://github.com/explosion/thinc",
     "note": "Functional deep learning library"
    },
    {
     "name": "spacy-models",
     "url": "https://github.com/explosion/spacy-models",
     "note": "Pretrained spaCy pipelines"
    }
   ],
   "openness": "Core libraries are released under the MIT licence.",
   "key_people": [
    {
     "name": "Ines Montani",
     "role": "CEO and founder",
     "url": "https://ines.io",
     "note": null
    },
    {
     "name": "Matthew Honnibal",
     "role": "CTO and founder",
     "url": "https://honnibal.dev",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact address printed on the website.",
   "sources": [
    "https://explosion.ai",
    "https://huggingface.co/explosion",
    "https://github.com/explosion"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "explosion",
   "mlat": 52.54165,
   "mlon": 13.46663,
   "logo": true
  },
  {
   "orig_id": "far-ai",
   "name": "FAR.AI",
   "alias": null,
   "full_name": "FAR.AI",
   "type": "nonprofit",
   "founded": 2022,
   "country": "United States",
   "city": "Berkeley",
   "lat": 37.8715,
   "lon": -122.273,
   "website": "https://far.ai",
   "huggingface": null,
   "github": "https://github.com/AlignmentResearch",
   "x": "https://x.com/FARAIResearch",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "AI safety research nonprofit publishing open code for adversarial robustness and interpretability research.",
   "tagline": "A nonprofit AI security research organization that builds safety solutions, tests frontier models and convenes the field.",
   "about": [
    "FAR.AI describes itself as an AI security research and adoption non-profit. It was co-founded by Adam Gleave, who is CEO, and Karl Berzins, who is president.",
    "Its work spans in-house research agendas, applied research that validates ideas through near-frontier-scale prototypes, pre-deployment testing and consulting with frontier developers, government capacity building, field-building events and grantmaking, and public safety evaluations. Its GitHub organisation, AlignmentResearch, publishes code from this work."
   ],
   "mission": "FAR.AI aims to close the gap between safety research and real-world adoption so that advanced AI is trustworthy, secure and beneficial.",
   "notice": null,
   "known_for": [
    "AI safety and security research",
    "Pre-deployment testing of frontier models",
    "Convenings and field-building events"
   ],
   "focus": [
    "AI safety",
    "adversarial robustness",
    "interpretability"
   ],
   "contributions": [],
   "github_highlights": [
    {
     "name": "tuned-lens",
     "url": "https://github.com/AlignmentResearch/tuned-lens",
     "note": "Tuned lens tools"
    },
    {
     "name": "go_attack",
     "url": "https://github.com/AlignmentResearch/go_attack",
     "note": "Adversarial attacks on Go AIs"
    },
    {
     "name": "vlmrm",
     "url": "https://github.com/AlignmentResearch/vlmrm",
     "note": "Vision-language models as rewards"
    },
    {
     "name": "learned-planner",
     "url": "https://github.com/AlignmentResearch/learned-planner",
     "note": "Learned planner research"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Eliciting Latent Predictions from Transformers with the Tuned Lens",
     "url": "https://arxiv.org/abs/2303.08112",
     "year": "2023"
    },
    {
     "title": "Adversarial Policies Beat Superhuman Go AIs",
     "url": "https://arxiv.org/abs/2211.00241",
     "year": "2022"
    },
    {
     "title": "Can Go AIs be adversarially robust?",
     "url": "https://arxiv.org/abs/2406.12843",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "tuned-lens",
     "url": "https://github.com/AlignmentResearch/tuned-lens",
     "note": "Tuned lens tools"
    },
    {
     "name": "go_attack",
     "url": "https://github.com/AlignmentResearch/go_attack",
     "note": "Adversarial attacks on Go AIs"
    },
    {
     "name": "vlmrm",
     "url": "https://github.com/AlignmentResearch/vlmrm",
     "note": "Vision-language models as rewards"
    }
   ],
   "openness": "MIT and Apache-2.0 research code on GitHub.",
   "key_people": [
    {
     "name": "Adam Gleave",
     "role": "Co-founder and CEO",
     "url": "https://far.ai/author/adam-gleave",
     "note": null
    },
    {
     "name": "Karl Berzins",
     "role": "Co-founder and President",
     "url": "https://far.ai/author/karl-berzins",
     "note": null
    },
    {
     "name": "Kellin Pelrine",
     "role": "Head of Technology, Frontier Safety",
     "url": "https://far.ai/author/kellin-pelrine",
     "note": null
    },
    {
     "name": "Boston Nyer",
     "role": "Chief Operating Officer",
     "url": "https://far.ai/author/boston-nyer",
     "note": null
    },
    {
     "name": "Chris Cundy",
     "role": "Member of Technical Staff",
     "url": "https://far.ai/author/chris-cundy",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.far.ai/",
    "https://github.com/AlignmentResearch"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "far-ai",
   "mlat": 37.85423,
   "mlon": -122.20565,
   "logo": true
  },
  {
   "orig_id": "fast-ai",
   "name": "fast.ai",
   "alias": null,
   "full_name": "fast.ai",
   "type": "institute",
   "founded": 2016,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.fast.ai",
   "huggingface": null,
   "github": "https://github.com/fastai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Research and education lab behind the fastai library and free deep learning courses.",
   "tagline": "A nonprofit research and education group, founded by Jeremy Howard and Rachel Thomas, behind the fastai library and free deep learning courses.",
   "about": [
    "fast.ai was founded in October 2016 by Jeremy Howard and Rachel Thomas with the goal of democratizing deep learning. Sylvain Gugger is also listed as a key person.",
    "Its free course Practical Deep Learning for Coders was first taught at the University of San Francisco and is delivered online with Python as the only prerequisite. The fastai library, built on PyTorch and licensed under Apache 2.0, reached version 1.0 in the fall of 2018. In 2018 fast.ai students won Stanford's DAWNBench CIFAR-10 challenge for fastest and cheapest training."
   ],
   "mission": "fast.ai works to make deep learning accessible to people from varied backgrounds, through free courses and open-source software.",
   "notice": null,
   "known_for": [
    "Practical Deep Learning for Coders",
    "The fastai library",
    "DAWNBench CIFAR-10 result"
   ],
   "focus": [
    "education",
    "libraries",
    "deep learning"
   ],
   "contributions": [
    {
     "title": "fast.ai founded",
     "note": "Jeremy Howard and Rachel Thomas established the group in October 2016.",
     "year": "2016"
    },
    {
     "title": "DAWNBench win",
     "note": "Students won the CIFAR-10 category of Stanford's DAWNBench for speed and cost.",
     "year": "2018"
    },
    {
     "title": "fastai 1.0",
     "note": "Version 1.0 of the open-source library on PyTorch was released.",
     "year": "2018"
    }
   ],
   "github_highlights": [
    {
     "name": "fastai",
     "url": "https://github.com/fastai/fastai",
     "note": "The fastai deep learning library."
    },
    {
     "name": "fastbook",
     "url": "https://github.com/fastai/fastbook",
     "note": "The fastai book as notebooks."
    },
    {
     "name": "numerical-linear-algebra",
     "url": "https://github.com/fastai/numerical-linear-algebra",
     "note": "Computational linear algebra course notebooks."
    },
    {
     "name": "courses",
     "url": "https://github.com/fastai/courses",
     "note": "fast.ai course materials."
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Universal Language Model Fine-tuning for Text Classification",
     "url": "https://arxiv.org/abs/1801.06146",
     "year": "2018"
    },
    {
     "title": "fastai: A Layered API for Deep Learning",
     "url": "https://arxiv.org/abs/2002.04688",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "fastai",
     "url": "https://github.com/fastai/fastai",
     "note": "Library"
    },
    {
     "name": "fastbook",
     "url": "https://github.com/fastai/fastbook",
     "note": "Book"
    }
   ],
   "openness": "Open-source Apache-2.0 code and free courses.",
   "key_people": [
    {
     "name": "Jeremy Howard",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Fast.ai",
     "note": null
    },
    {
     "name": "Rachel Thomas",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Fast.ai",
     "note": null
    },
    {
     "name": "Sylvain Gugger",
     "role": "Key contributor",
     "url": "https://en.wikipedia.org/wiki/Fast.ai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.fast.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.fast.ai",
    "https://github.com/fastai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "fast-ai",
   "mlat": 37.86202,
   "mlon": -122.315,
   "logo": true
  },
  {
   "orig_id": "fbk-mt",
   "name": "FBK Machine Translation Unit",
   "alias": null,
   "full_name": "FBK Machine Translation Unit",
   "type": "institute",
   "founded": null,
   "country": "Italy",
   "city": "Trento",
   "lat": 46.0748,
   "lon": 11.1217,
   "website": "https://mt.fbk.eu",
   "huggingface": "https://huggingface.co/FBK-MT",
   "github": "https://github.com/hlt-mt",
   "x": "https://x.com/fbk_mt",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Machine translation and speech research unit at Fondazione Bruno Kessler that publishes open speech and translation datasets and code.",
   "tagline": "The machine translation group at Fondazione Bruno Kessler, publishing open speech and text translation datasets and software.",
   "about": [
    "The MT group at Fondazione Bruno Kessler works on speech and text technologies for multilingual communication. Its research covers neural architectures and multimodal resources for applications such as subtitling, interpreting and gender inclusivity, and it sits within FBK's Digital Industry center.",
    "The group's site lists corpora including MCIF, mGeNTE, MOSEL, Speech-MASSIVE and WAGS, and software such as simulstream. Researchers named on the site include Luisa Bentivogli and Beatrice Savoldi."
   ],
   "mission": "The unit aims to advance speech and text translation, with attention to inclusivity, and to share its data and code openly.",
   "notice": null,
   "known_for": [
    "Open speech translation corpora",
    "Simultaneous translation software",
    "Gender-fair machine translation research"
   ],
   "focus": [
    "speech translation",
    "speech",
    "datasets"
   ],
   "contributions": [
    {
     "title": "simulstream release",
     "note": "Software for simultaneous speech translation released.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "mosel",
     "url": "https://github.com/hlt-mt/mosel",
     "note": "Collection of open source speech data (166 stars)"
    },
    {
     "name": "FBK-fairseq",
     "url": "https://github.com/hlt-mt/FBK-fairseq",
     "note": "Open source code of works from the FBK MT unit (60 stars)"
    },
    {
     "name": "TMOP",
     "url": "https://github.com/hlt-mt/TMOP",
     "note": "Translation Memory Open-source Purifier (35 stars)"
    },
    {
     "name": "simulstream",
     "url": "https://github.com/hlt-mt/simulstream",
     "note": "Library for simultaneous speech recognition and translation (33 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "FBK-MT/mosel",
     "url": "https://huggingface.co/datasets/FBK-MT/mosel",
     "kind": "dataset",
     "note": "MOSEL speech data collection, the unit's most downloaded repository."
    },
    {
     "name": "FBK-MT/Speech-MASSIVE",
     "url": "https://huggingface.co/datasets/FBK-MT/Speech-MASSIVE",
     "kind": "dataset",
     "note": "Multilingual spoken language understanding dataset."
    },
    {
     "name": "FBK-MT/MCIF",
     "url": "https://huggingface.co/datasets/FBK-MT/MCIF",
     "kind": "dataset",
     "note": "Crosslingual instruction-following benchmark."
    },
    {
     "name": "FBK-MT/fama-medium",
     "url": "https://huggingface.co/FBK-MT/fama-medium",
     "kind": "model",
     "note": "Open-science speech foundation model."
    }
   ],
   "papers": [
    {
     "title": "MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages",
     "url": "https://arxiv.org/abs/2410.01036",
     "year": "2024"
    },
    {
     "title": "Speech-MASSIVE: A Multilingual Speech Dataset for SLU and Beyond",
     "url": "https://arxiv.org/abs/2408.03900",
     "year": "2024"
    },
    {
     "title": "FAMA: The First Large-Scale Open-Science Speech Foundation Model for English and Italian",
     "url": "https://arxiv.org/abs/2505.22759",
     "year": "2025"
    },
    {
     "title": "MCIF: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks",
     "url": "https://arxiv.org/abs/2507.19634",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "MOSEL",
     "url": "https://huggingface.co/datasets/FBK-MT/mosel",
     "note": "Open-source speech data collection"
    },
    {
     "name": "FAMA models",
     "url": "https://huggingface.co/FBK-MT/fama-medium",
     "note": "Open ASR models"
    },
    {
     "name": "FBK-fairseq",
     "url": "https://github.com/hlt-mt/FBK-fairseq",
     "note": "Code for MT unit works"
    },
    {
     "name": "simulstream",
     "url": "https://github.com/hlt-mt/simulstream",
     "note": "Library for streaming speech recognition and translation"
    }
   ],
   "openness": "Code and datasets are public on GitHub and Hugging Face under Apache-2.0 for several repositories.",
   "key_people": [
    {
     "name": "Luisa Bentivogli",
     "role": "Head of Unit",
     "url": "https://mt.fbk.eu/people/",
     "note": null
    },
    {
     "name": "Matteo Negri",
     "role": "Senior Research Scientist",
     "url": "https://mt.fbk.eu/people/",
     "note": null
    },
    {
     "name": "Mauro Cettolo",
     "role": "Senior Research Scientist",
     "url": "https://mt.fbk.eu/people/",
     "note": null
    },
    {
     "name": "Marco Gaido",
     "role": "Researcher",
     "url": "https://mt.fbk.eu/people/",
     "note": null
    },
    {
     "name": "Beatrice Savoldi",
     "role": "Researcher, gender bias in NLP",
     "url": "https://mt.fbk.eu/people/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The unit's website blocked automated access during checking; use the GitHub (hlt-mt) or Hugging Face pages for project contact.",
   "sources": [
    "https://mt.fbk.eu",
    "https://huggingface.co/FBK-MT",
    "https://github.com/hlt-mt"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "fbk-machine-translation-unit",
   "logo": true
  },
  {
   "orig_id": "flowers-inria",
   "name": "Flowers",
   "alias": "Inria and Univ. Bordeaux",
   "full_name": "Flowers (Inria and Univ. Bordeaux)",
   "type": "institute",
   "founded": null,
   "country": "France",
   "city": "Bordeaux",
   "lat": 44.8378,
   "lon": -0.5792,
   "website": "https://flowers.inria.fr",
   "huggingface": null,
   "github": "https://github.com/flowersteam",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Inria research team on developmental and curiosity-driven AI, open-ended learning and grounding language models through reinforcement learning.",
   "tagline": "Flowers is an Inria team in Bordeaux studying curiosity-driven learning in humans and machines, led by Pierre-Yves Oudeyer.",
   "about": [
    "The Flowers team investigates the mechanisms of sensorimotor, cognitive and social development in humans and machines. It is housed at Inria Bordeaux-Sud-Ouest and Univ. Bordeaux and is headed by Pierre-Yves Oudeyer, a research director known for work on artificial curiosity and intrinsic motivation.",
    "The team's GitHub organization, opened in 2012, publishes research code such as the Explauto exploration library and testbeds for automatic curriculum learning in deep reinforcement learning. More recent releases ground language models in interactive environments using online reinforcement learning and provide the Lamorel library for that purpose."
   ],
   "mission": "To understand how open-ended, curiosity-driven learning works and to build artificial agents that develop and learn in similar ways.",
   "notice": null,
   "known_for": [
    "Curiosity-driven and developmental AI",
    "Pierre-Yves Oudeyer's research team",
    "Grounding LLMs with reinforcement learning"
   ],
   "focus": [
    "reinforcement learning",
    "open-ended learning",
    "LLM agents",
    "robotics"
   ],
   "contributions": [
    {
     "title": "GitHub organization opened",
     "note": "The team began publishing research code under flowersteam.",
     "year": "2012"
    },
    {
     "title": "Explauto library",
     "note": "An autonomous exploration library for developmental robotics was released.",
     "year": "2014"
    },
    {
     "title": "teachDeepRL",
     "note": "Code for teacher algorithms in deep reinforcement learning was published.",
     "year": "2019"
    },
    {
     "title": "TeachMyAgent",
     "note": "A testbed platform for automatic curriculum learning methods in deep RL was released.",
     "year": "2021"
    },
    {
     "title": "Grounding LLMs with online RL and Lamorel",
     "note": "Code for functionally grounding LLM knowledge in BabyAI-Text and a library for RL with LLMs were published.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "Grounding_LLMs_with_online_RL",
     "url": "https://github.com/flowersteam/Grounding_LLMs_with_online_RL",
     "note": "We perform functional grounding of LLMs' knowledge in BabyAI-Text (277 stars)"
    },
    {
     "name": "lamorel",
     "url": "https://github.com/flowersteam/lamorel",
     "note": "Lamorel is a Python library designed for RL practitioners eager to use Large Language Models (LLMs). (247 stars)"
    },
    {
     "name": "teachDeepRL",
     "url": "https://github.com/flowersteam/teachDeepRL",
     "note": "Open-source repository. (94 stars)"
    },
    {
     "name": "TeachMyAgent",
     "url": "https://github.com/flowersteam/TeachMyAgent",
     "note": "TeachMyAgent is a testbed platform for Automatic Curriculum Learning methods in Deep RL. (78 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Grounding LLMs with online RL",
     "url": "https://github.com/flowersteam/Grounding_LLMs_with_online_RL",
     "note": "Functional grounding of LLMs in BabyAI-Text"
    },
    {
     "name": "lamorel",
     "url": "https://github.com/flowersteam/lamorel",
     "note": "Library for using LLMs in RL training loops"
    },
    {
     "name": "TeachMyAgent",
     "url": "https://github.com/flowersteam/TeachMyAgent",
     "note": "Testbed for automatic curriculum learning in deep RL"
    },
    {
     "name": "teachDeepRL",
     "url": "https://github.com/flowersteam/teachDeepRL",
     "note": "Teacher algorithms for deep RL"
    }
   ],
   "openness": "Research code is released openly in the flowersteam GitHub organisation, typically alongside papers.",
   "key_people": [
    {
     "name": "Pierre-Yves Oudeyer",
     "role": "Team head, Inria research director",
     "url": "https://en.wikipedia.org/wiki/Pierre-Yves_Oudeyer",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the team website.",
   "sources": [
    "https://flowers.inria.fr",
    "https://github.com/flowersteam"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "flowers",
   "logo": true
  },
  {
   "orig_id": "futurehouse",
   "name": "FutureHouse",
   "alias": null,
   "full_name": "FutureHouse",
   "type": "nonprofit",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.futurehouse.org",
   "huggingface": "https://huggingface.co/futurehouse",
   "github": "https://github.com/Future-House",
   "x": "https://x.com/futurehouse",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Nonprofit AI-for-science lab building language agents and open models for scientific discovery.",
   "tagline": "A nonprofit AI-for-science lab, founded by Sam Rodriques and Andrew White, building language agents to automate scientific discovery.",
   "about": [
    "FutureHouse was founded in 2023 by Sam Rodriques and Andrew White as a 501(c)(3) nonprofit funded by philanthropy and grants. Eric and Wendy Schmidt funded its first two years, with later support from Open Philanthropy, the National Science Foundation and the AI Safety Institute, among others.",
    "Its stated mission is to build an AI Scientist that automates and accelerates fundamental science to improve human health. Its work includes PaperQA2 for scientific literature, the Aviary framework for training language agents, the Robin multi-agent system, the ether0 chemistry reasoning model and the LAB-Bench and BixBench benchmarks. Systems originally developed under the FutureHouse name are now part of Edison Scientific's platform."
   ],
   "mission": "FutureHouse aims to build an AI Scientist that speeds up fundamental research relevant to human health.",
   "notice": null,
   "known_for": [
    "PaperQA2 literature agent",
    "Robin hypothesis-generation system",
    "ether0 chemistry model",
    "LAB-Bench and BixBench benchmarks"
   ],
   "focus": [
    "AI for science",
    "agents",
    "chemistry",
    "biology benchmarks"
   ],
   "contributions": [
    {
     "title": "FutureHouse founded",
     "note": "Founded in San Francisco as a nonprofit by Sam Rodriques and Andrew White.",
     "year": "2023"
    },
    {
     "title": "Edison Scientific spin-out",
     "note": "A for-profit company, Edison Scientific, spun out of FutureHouse.",
     "year": "2025"
    },
    {
     "title": "Kosmos",
     "note": "Launched an AI Scientist for autonomous complex analyses.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "paper-qa",
     "url": "https://github.com/Future-House/paper-qa",
     "note": "RAG for scientific documents"
    },
    {
     "name": "ether0",
     "url": "https://github.com/Future-House/ether0",
     "note": "Scientific reasoning model"
    },
    {
     "name": "aviary",
     "url": "https://github.com/Future-House/aviary",
     "note": "Language agent gym"
    },
    {
     "name": "robin",
     "url": "https://github.com/Future-House/robin",
     "note": "Multi-agent scientific discovery system"
    }
   ],
   "hf_highlights": [
    {
     "name": "ether0",
     "url": "https://huggingface.co/futurehouse/ether0",
     "kind": "model",
     "note": "Chemistry reasoning model"
    },
    {
     "name": "BixBench",
     "url": "https://huggingface.co/datasets/futurehouse/BixBench",
     "kind": "dataset",
     "note": "Bioinformatics benchmark"
    },
    {
     "name": "lab-bench",
     "url": "https://huggingface.co/datasets/futurehouse/lab-bench",
     "kind": "dataset",
     "note": "Biology research benchmark"
    }
   ],
   "papers": [
    {
     "title": "Language agents achieve superhuman synthesis of scientific knowledge",
     "url": "https://arxiv.org/abs/2409.13740",
     "year": "2024"
    },
    {
     "title": "LAB-Bench: Measuring Capabilities of Language Models for Biology Research",
     "url": "https://arxiv.org/abs/2407.10362",
     "year": "2024"
    },
    {
     "title": "Training a Scientific Reasoning Model for Chemistry",
     "url": "https://arxiv.org/abs/2506.17238",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "paper-qa",
     "url": "https://github.com/Future-House/paper-qa",
     "note": "RAG for scientific documents"
    },
    {
     "name": "ether0",
     "url": "https://github.com/Future-House/ether0",
     "note": "Scientific reasoning model"
    },
    {
     "name": "aviary",
     "url": "https://github.com/Future-House/aviary",
     "note": "Language agent gym"
    }
   ],
   "openness": "Apache-2.0 code; open ether0 weights and benchmark datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Sam Rodriques",
     "role": "CEO and co-founder",
     "url": "https://www.futurehouse.org/team/sam-rodriques",
     "note": null
    },
    {
     "name": "Andrew White",
     "role": "Head of Science and co-founder",
     "url": "https://www.futurehouse.org/team/andrew-white",
     "note": null
    },
    {
     "name": "Jessica Wu",
     "role": "Head of Operations",
     "url": "https://www.futurehouse.org/team/jessica-wu",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.futurehouse.org",
    "https://github.com/Future-House",
    "https://huggingface.co/futurehouse"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "futurehouse",
   "mlat": 37.87772,
   "mlon": -122.34113,
   "logo": true
  },
  {
   "orig_id": "genmo",
   "name": "Genmo",
   "alias": null,
   "full_name": "Genmo",
   "type": "startup",
   "founded": null,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.genmo.ai",
   "huggingface": "https://huggingface.co/genmo",
   "github": "https://github.com/genmoai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "hi@genmo.ai",
   "summary": "Video generation lab that released Mochi 1 as an open model.",
   "tagline": "A video generation company that released Mochi 1, an open-source video model under the Apache 2.0 license.",
   "about": [
    "Genmo is a US video generation company whose first public open-source release was Mochi 1, licensed under Apache 2.0 for individual and commercial use. Its team describes itself as including original creators of foundational AI research.",
    "The company lists investors and advisors including Ion Stoica, Pieter Abbeel, Joseph Gonzalez, Amjad Masad and Michele Catasta. Weights and code are published on Hugging Face and GitHub."
   ],
   "mission": "Genmo aims to advance video generation and to share an open model with the research and developer community.",
   "notice": null,
   "known_for": [
    "Mochi 1 open video model",
    "Apache 2.0 licensed release"
   ],
   "focus": [
    "video generation"
   ],
   "contributions": [
    {
     "title": "Mochi 1",
     "note": "First public open-source video model, released under Apache 2.0.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "mochi",
     "url": "https://github.com/genmoai/mochi",
     "note": "Open video generation models."
    }
   ],
   "hf_highlights": [
    {
     "name": "mochi-1-preview",
     "url": "https://huggingface.co/genmo/mochi-1-preview",
     "kind": "model",
     "note": "Mochi 1 preview weights."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "mochi",
     "url": "https://github.com/genmoai/mochi",
     "note": "Code"
    },
    {
     "name": "mochi-1-preview",
     "url": "https://huggingface.co/genmo/mochi-1-preview",
     "note": "Weights"
    }
   ],
   "openness": "Open weights (Apache-2.0).",
   "key_people": [
    {
     "name": "Paras Jain",
     "role": "Co-founder and CEO",
     "url": "https://www.forwardfuture.ai/p/reimagining-creativity-paras-jain-on-genmo-s-vision-and-the-future-of-generative-ai",
     "note": null
    },
    {
     "name": "Ajay Jain",
     "role": "Co-founder",
     "url": "https://thursdai.news/guests/ajayj_",
     "note": "Sources differ on his title."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.genmo.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.genmo.ai",
    "https://huggingface.co/genmo"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "genmo",
   "mlat": 37.88862,
   "mlon": -122.37092,
   "logo": true
  },
  {
   "orig_id": "gensyn",
   "name": "Gensyn",
   "alias": null,
   "full_name": "Gensyn",
   "type": "startup",
   "founded": 2020,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://www.gensyn.ai",
   "huggingface": null,
   "github": "https://github.com/gensyn-ai",
   "x": "https://x.com/gensynai",
   "discord": "https://discord.com/invite/gensyn",
   "linkedin": "https://www.linkedin.com/company/gensynai",
   "contact_email": null,
   "summary": "London-based machine learning research and production lab building decentralised training and verification infrastructure. It open-sources frameworks such as RL Swarm and also works on tools like REE and Delphi information markets.",
   "tagline": "A London-based project building decentralised infrastructure for machine learning, founded by Ben Fielding and Harry Grieve.",
   "about": [
    "Gensyn was founded in 2020 in London by Ben Fielding and Harry Grieve. It describes its mission as building machines that predict the future and works on predictive intelligence infrastructure and verifiable AI.",
    "The company has raised around 50 million dollars, including a 43 million dollar Series A led by a16z crypto in 2023, with backing from CoinFund, Protocol Labs, Canonical Crypto, Eden Block and Maven 11. Its GitHub organisation publishes open-source frameworks, including RL Swarm for collaborative training over the internet."
   ],
   "mission": "Gensyn aims to make machine learning compute open and verifiable by coordinating training across a decentralised network.",
   "notice": null,
   "known_for": [
    "Decentralised machine learning infrastructure",
    "RL Swarm collaborative training",
    "Verifiable AI research"
   ],
   "focus": [
    "decentralised training",
    "reinforcement learning",
    "infrastructure"
   ],
   "contributions": [
    {
     "title": "Gensyn founded",
     "note": "Founded in London by Ben Fielding and Harry Grieve.",
     "year": "2020"
    },
    {
     "title": "Series A",
     "note": "a16z crypto led a $43 million Series A.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "rl-swarm",
     "url": "https://github.com/gensyn-ai/rl-swarm",
     "note": "Open framework for RL training swarms over the internet (1,672 stars)"
    },
    {
     "name": "codeassist",
     "url": "https://github.com/gensyn-ai/codeassist",
     "note": "Private local AI coding assistant (699 stars)"
    },
    {
     "name": "blockassist",
     "url": "https://github.com/gensyn-ai/blockassist",
     "note": "Assistant that learns from a player's Minecraft actions (102 stars)"
    },
    {
     "name": "skippipe",
     "url": "https://github.com/gensyn-ai/skippipe",
     "note": "Code for the SkipPipe pipelining paper (49 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "SkipPipe: Partial and Reordered Pipelining Framework for Training LLMs in Heterogeneous Networks",
     "url": "https://arxiv.org/abs/2502.19913",
     "year": "2025"
    },
    {
     "title": "Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing",
     "url": "https://arxiv.org/abs/2509.08721",
     "year": "2025"
    },
    {
     "title": "Verde: Verification via Refereed Delegation for Machine Learning Programs",
     "url": "https://arxiv.org/abs/2502.19405",
     "year": "2025"
    },
    {
     "title": "NoLoCo: No-all-reduce Low Communication Training Method for Large Models",
     "url": "https://arxiv.org/abs/2506.10911",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "RL Swarm",
     "url": "https://github.com/gensyn-ai/rl-swarm",
     "note": "Open framework for RL training swarms over the internet"
    },
    {
     "name": "CodeAssist",
     "url": "https://github.com/gensyn-ai/codeassist",
     "note": "Local private coding assistant"
    },
    {
     "name": "SkipPipe",
     "url": "https://github.com/gensyn-ai/skippipe",
     "note": "Code for a pipelining paper"
    }
   ],
   "openness": "Main repositories are open source, mostly under MIT.",
   "key_people": [
    {
     "name": "Ben Fielding",
     "role": "Co-founder",
     "url": "https://www.gensyn.ai/about",
     "note": null
    },
    {
     "name": "Harry Grieve",
     "role": "Co-founder",
     "url": "https://www.gensyn.ai/about",
     "note": null
    }
   ],
   "discord_note": "Public community server linked from the Gensyn site for discussion and announcements.",
   "contact_note": "Gensyn's site links its research, news and community Discord; no organisation email is printed on the home page.",
   "sources": [
    "https://www.gensyn.ai",
    "https://github.com/gensyn-ai"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "gensyn",
   "mlat": 51.52212,
   "mlon": -0.02419,
   "logo": true
  },
  {
   "orig_id": "ghana-nlp",
   "name": "Ghana NLP",
   "alias": null,
   "full_name": "Ghana NLP",
   "type": "collective",
   "founded": null,
   "country": "Ghana",
   "city": "Accra",
   "lat": 5.6037,
   "lon": -0.187,
   "website": "https://ghananlp.org",
   "huggingface": "https://huggingface.co/Ghana-NLP",
   "github": "https://github.com/GhanaNLP",
   "x": "https://twitter.com/GhanaNLP",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/ghana-nlp/",
   "contact_email": null,
   "summary": "Ghana NLP is an open-source initiative working on NLP for Ghanaian languages and applications to local problems. It publishes Twi and other language models and parallel text datasets on Hugging Face and code on GitHub.",
   "tagline": "An open-source community building language technology for Ghanaian languages such as Twi, Ewe and Frafra.",
   "about": [
    "Ghana NLP is an open-source initiative working on natural language processing for Ghanaian languages and its application to local problems. It aims to build better training data, adapt state-of-the-art techniques to low-resource settings and create working systems such as translators.",
    "Its projects include the Khaya Translator apps, Ghana Speech ID, the Nsanku translation and ASR benchmarks, and ABENA. The community reports about 600 members and invites volunteers through a join form."
   ],
   "mission": "The initiative works to bring Ghanaian languages into modern language technology, beginning with Twi and extending to others.",
   "notice": null,
   "known_for": [
    "Khaya translation apps",
    "Twi language resources",
    "Nsanku benchmarks"
   ],
   "focus": [
    "low-resource languages",
    "language models",
    "datasets",
    "machine translation"
   ],
   "contributions": [
    {
     "title": "ABENA",
     "note": "Introduced a BERT-based language model for Twi.",
     "year": "2020"
    },
    {
     "title": "ghana-speech dataset",
     "note": "Released 2,247 hours of speech across 42 Ghanaian languages, followed by a phoneme ASR model.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "kasa",
     "url": "https://github.com/GhanaNLP/kasa",
     "note": "English to Twi translation system."
    },
    {
     "name": "ABENA",
     "url": "https://github.com/GhanaNLP/ABENA",
     "note": "Training scripts for language models."
    },
    {
     "name": "nsanku",
     "url": "https://github.com/GhanaNLP/nsanku",
     "note": "Zero-shot LLM translation eval."
    },
    {
     "name": "ghanaian-nlp-datasets-models",
     "url": "https://github.com/GhanaNLP/ghanaian-nlp-datasets-models",
     "note": "Curated datasets and models."
    }
   ],
   "hf_highlights": [
    {
     "name": "distilabena-base-v2-akuapem-twi-cased",
     "url": "https://huggingface.co/Ghana-NLP/distilabena-base-v2-akuapem-twi-cased",
     "kind": "model",
     "note": "DistilABENA."
    },
    {
     "name": "abena-base-asante-twi-uncased",
     "url": "https://huggingface.co/Ghana-NLP/abena-base-asante-twi-uncased",
     "kind": "model",
     "note": "ABENA Asante Twi."
    },
    {
     "name": "TWI_ENGLISH_PARALLEL_TEXT",
     "url": "https://huggingface.co/datasets/Ghana-NLP/TWI_ENGLISH_PARALLEL_TEXT",
     "kind": "dataset",
     "note": "Twi-English."
    },
    {
     "name": "EWE_ENGLISH_PARALLEL_TEXT",
     "url": "https://huggingface.co/datasets/Ghana-NLP/EWE_ENGLISH_PARALLEL_TEXT",
     "kind": "dataset",
     "note": "Ewe-English."
    }
   ],
   "papers": [
    {
     "title": "Nsanku: Evaluating Zero-Shot Translation Performance of LLMs for Ghanaian Languages",
     "url": "https://arxiv.org/abs/2605.04208",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "ABENA Twi BERT models",
     "url": "https://huggingface.co/Ghana-NLP/abena-base-asante-twi-uncased",
     "note": "Pretrained Twi language models"
    },
    {
     "name": "Twi-English parallel text",
     "url": "https://huggingface.co/datasets/Ghana-NLP/TWI_ENGLISH_PARALLEL_TEXT",
     "note": "Parallel text dataset"
    },
    {
     "name": "Ewe-English parallel text",
     "url": "https://huggingface.co/datasets/Ghana-NLP/EWE_ENGLISH_PARALLEL_TEXT",
     "note": "Parallel text dataset"
    },
    {
     "name": "kasa",
     "url": "https://github.com/GhanaNLP/kasa",
     "note": "English to Twi translation system"
    }
   ],
   "openness": "Models and datasets are published on Hugging Face and code on GitHub; licences differ per resource.",
   "key_people": [
    {
     "name": "Paul Azunre",
     "role": "Co-founder",
     "url": "https://yen.com.gh/education/255014-paul-azunre-nsmq-winner-develops-app-translates-twi-english/",
     "note": "Described as co-founder of Ghana NLP in press coverage."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://ghananlp.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://ghananlp.org",
    "https://huggingface.co/Ghana-NLP",
    "https://github.com/GhanaNLP"
   ],
   "confidence": "high",
   "region": "Africa",
   "id": "ghana-nlp",
   "logo": true
  },
  {
   "orig_id": "giskard",
   "name": "Giskard",
   "alias": null,
   "full_name": "Giskard",
   "type": "startup",
   "founded": 2021,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.giskard.ai",
   "huggingface": "https://huggingface.co/giskardai",
   "github": "https://github.com/Giskard-AI",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/giskard-ai",
   "contact_email": null,
   "summary": "Paris company building evaluation and red-teaming tools for AI agents and LLMs. It maintains an open-source testing library and publishes safety benchmarks.",
   "tagline": "A Paris company building AI red-teaming and evaluation tools, with an open-source testing library and the Phare benchmark.",
   "about": [
    "Giskard describes its mission as securing conversational AI, with a platform for red teaming, LLM evaluation and vulnerability scanning. Alex Combessie and Jean-Marie John-Mathews are co-founders and co-CEOs.",
    "The company maintains the open-source Giskard library and publishes the Phare benchmark. Its partners include the European Commission, Bpifrance, Hugging Face and Mistral AI, and investors include Julien Chaumond and Elaia."
   ],
   "mission": "Giskard aims to make AI agents and language models safe and reliable through systematic testing.",
   "notice": null,
   "known_for": [
    "Open-source LLM testing library",
    "Phare safety benchmark",
    "AI red teaming tools"
   ],
   "focus": [
    "evaluation",
    "safety",
    "red-teaming",
    "tooling"
   ],
   "contributions": [],
   "github_highlights": [
    {
     "name": "giskard-oss",
     "url": "https://github.com/Giskard-AI/giskard-oss",
     "note": "🐢 Open-Source Evaluation & Testing library for LLM Agents (5,956 stars)"
    },
    {
     "name": "awesome-ai-safety",
     "url": "https://github.com/Giskard-AI/awesome-ai-safety",
     "note": "📚 A curated list of papers & technical articles on AI Quality & Safety (228 stars)"
    },
    {
     "name": "prompt-injections",
     "url": "https://github.com/Giskard-AI/prompt-injections",
     "note": "💉 Collection of prompt injections used in the open-source Giskard Scan (38 stars)"
    },
    {
     "name": "giskard-vision",
     "url": "https://github.com/Giskard-AI/giskard-vision",
     "note": "📸  Open-Source Evaluation & Testing for Computer Vision AI systems (31 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "giskardai/StereoTales",
     "url": "https://huggingface.co/datasets/giskardai/StereoTales",
     "kind": "dataset",
     "note": "Dataset: 387 downloads, 2 likes."
    },
    {
     "name": "giskardai/giskard-evaluator",
     "url": "https://huggingface.co/spaces/giskardai/giskard-evaluator",
     "kind": "space",
     "note": "Space: 5 likes on the Hub."
    },
    {
     "name": "giskardai/harmbench-scenarios",
     "url": "https://huggingface.co/datasets/giskardai/harmbench-scenarios",
     "kind": "dataset",
     "note": "Dataset: 325 downloads, 0 likes."
    },
    {
     "name": "giskardai/phare",
     "url": "https://huggingface.co/datasets/giskardai/phare",
     "kind": "dataset",
     "note": "Dataset: 272 downloads, 14 likes."
    }
   ],
   "papers": [
    {
     "title": "Phare: A Safety Probe for Large Language Models",
     "url": "https://arxiv.org/abs/2505.11365",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "giskard-oss",
     "url": "https://github.com/Giskard-AI/giskard-oss",
     "note": "Open-source evaluation and testing library for LLM agents"
    },
    {
     "name": "Phare",
     "url": "https://huggingface.co/datasets/giskardai/phare",
     "note": "LLM safety benchmark dataset"
    },
    {
     "name": "RealHarm",
     "url": "https://huggingface.co/datasets/giskardai/realharm",
     "note": "Dataset of real-world harm incidents"
    }
   ],
   "openness": "Core library is Apache-2.0 on GitHub; benchmarks are public on Hugging Face.",
   "key_people": [
    {
     "name": "Alex Combessie",
     "role": "Co-founder and co-CEO",
     "url": "https://www.giskard.ai/about",
     "note": null
    },
    {
     "name": "Jean-Marie John-Mathews",
     "role": "Co-founder and co-CEO",
     "url": "https://www.giskard.ai/about",
     "note": null
    },
    {
     "name": "Pierre Le Jeune",
     "role": "Head of AI Research",
     "url": "https://www.giskard.ai/about",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website.",
   "sources": [
    "https://www.giskard.ai",
    "https://huggingface.co/giskardai",
    "https://github.com/Giskard-AI"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "giskard",
   "mlat": 48.81065,
   "mlon": 2.49244,
   "logo": true
  },
  {
   "orig_id": "goodfire",
   "name": "Goodfire",
   "alias": null,
   "full_name": "Goodfire",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.goodfire.ai",
   "huggingface": "https://huggingface.co/Goodfire",
   "github": "https://github.com/goodfire-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Interpretability research lab releasing open sparse autoencoders and interpretability artefacts.",
   "tagline": "An interpretability research company, founded by researchers from OpenAI and Google DeepMind, that works to understand and align AI models.",
   "about": [
    "Goodfire says it was founded by researchers who helped pioneer interpretability at OpenAI and Google DeepMind. Its team includes founding members of those efforts, professors on leave and engineers, and it reports raising over 200 million dollars from investors including Menlo, Lightspeed and B Capital.",
    "Its stated mission is to advance the research and technology needed to understand and align AI so that people can trust superintelligence. Its Hugging Face and GitHub organisations host sparse autoencoders and other interpretability artifacts."
   ],
   "mission": "Goodfire aims to make the internal workings of AI models understandable so they can be aligned and trusted.",
   "notice": null,
   "known_for": [
    "Interpretability research",
    "Open sparse autoencoders",
    "Neurons as an interpretable basis"
   ],
   "focus": [
    "interpretability",
    "safety"
   ],
   "contributions": [
    {
     "title": "Goodfire founded",
     "note": "Founded in San Francisco by Eric Ho, Dan Balsam and Tom McGrath.",
     "year": "2024"
    },
    {
     "title": "Ember",
     "note": "Launched a platform for mechanistic interpretability research.",
     "year": "2024"
    },
    {
     "title": "Series A",
     "note": "Announced a $50 million Series A led by Menlo Ventures.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "r1-interpretability",
     "url": "https://github.com/goodfire-ai/r1-interpretability",
     "note": "R1 interpretability artefacts."
    },
    {
     "name": "param-decomp",
     "url": "https://github.com/goodfire-ai/param-decomp",
     "note": "Parameter decomposition."
    },
    {
     "name": "causalab",
     "url": "https://github.com/goodfire-ai/causalab",
     "note": "Causal abstraction tooling."
    },
    {
     "name": "goodfire-sdk",
     "url": "https://github.com/goodfire-ai/goodfire-sdk",
     "note": "Ember SDK."
    }
   ],
   "hf_highlights": [
    {
     "name": "Llama-3.1-8B-Instruct-SAE-l19",
     "url": "https://huggingface.co/Goodfire/Llama-3.1-8B-Instruct-SAE-l19",
     "kind": "model",
     "note": "SAE on Llama 3.1 8B."
    },
    {
     "name": "Llama-3.3-70B-Instruct-SAE-l50",
     "url": "https://huggingface.co/Goodfire/Llama-3.3-70B-Instruct-SAE-l50",
     "kind": "model",
     "note": "SAE on Llama 3.3 70B."
    },
    {
     "name": "reward-hack-data",
     "url": "https://huggingface.co/datasets/Goodfire/reward-hack-data",
     "kind": "dataset",
     "note": "Reward hacking data."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "r1-interpretability",
     "url": "https://github.com/goodfire-ai/r1-interpretability",
     "note": "R1 SAE artefacts"
    }
   ],
   "openness": "Open SAE weights and code.",
   "key_people": [
    {
     "name": "Eric Ho",
     "role": "Co-founder and CEO",
     "url": "https://sequoiacap.com/podcast/training-data-eric-ho",
     "note": null
    },
    {
     "name": "Tom McGrath",
     "role": "Co-founder and Chief Scientist",
     "url": "https://lsvp.com/stories/goodfire-building-interpretable-ai/",
     "note": "Previously co-founded Google DeepMind's interpretability team."
    },
    {
     "name": "Dan Balsam",
     "role": "Co-founder and CTO",
     "url": "https://lsvp.com/stories/goodfire-building-interpretable-ai/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.goodfire.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.goodfire.ai",
    "https://huggingface.co/Goodfire"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "goodfire",
   "mlat": 37.8942,
   "mlon": -122.40299,
   "logo": true
  },
  {
   "orig_id": "gorilla",
   "name": "Gorilla",
   "alias": "UC Berkeley",
   "full_name": "Gorilla (UC Berkeley)",
   "type": "academic",
   "founded": 2023,
   "country": "United States",
   "city": "Berkeley",
   "lat": 37.8715,
   "lon": -122.273,
   "website": "https://gorilla.cs.berkeley.edu",
   "huggingface": "https://huggingface.co/gorilla-llm",
   "github": "https://github.com/ShishirPatil",
   "x": null,
   "discord": "https://discord.gg/grXXvj9Whz",
   "linkedin": null,
   "contact_email": null,
   "summary": "UC Berkeley research project releasing open LLMs and the Berkeley Function-Calling Leaderboard for tool and API use.",
   "tagline": "A UC Berkeley project that built Gorilla LLM and the Berkeley Function-Calling Leaderboard for evaluating tool and API use by language models.",
   "about": [
    "Gorilla is a UC Berkeley research project led by Shishir G. Patil, Tianjun Zhang, Xin Wang and Joseph E. Gonzalez. It produced Gorilla LLM, an open-source model that invokes API calls, and a paper posted as arXiv 2305.15334.",
    "Related releases include OpenFunctions v2, a 6.91B-parameter function-calling model, RAFT for retrieval-aware fine-tuning, GoEX, a runtime for executing LLM-generated actions with undo and damage confinement, and the Berkeley Function-Calling Leaderboard, which covers about 2,000 question-function-answer pairs across Python, Java, JavaScript and REST."
   ],
   "mission": "Gorilla aims to let language models use tools and APIs reliably, and to give the community open models and benchmarks for measuring that ability.",
   "notice": null,
   "known_for": [
    "Gorilla LLM for API calls",
    "Berkeley Function-Calling Leaderboard",
    "RAFT fine-tuning method",
    "GoEX execution engine"
   ],
   "focus": [
    "function calling",
    "tool use",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Gorilla paper",
     "note": "Introduced an LLM connected to large API collections.",
     "year": "2023"
    },
    {
     "title": "RAFT",
     "note": "Retrieval-aware fine-tuning method.",
     "year": "2024"
    },
    {
     "title": "GoEX",
     "note": "Runtime for LLM-generated actions.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "gorilla",
     "url": "https://github.com/ShishirPatil/gorilla",
     "note": "Training and evaluating LLMs for function calls"
    }
   ],
   "hf_highlights": [
    {
     "name": "Berkeley-Function-Calling-Leaderboard",
     "url": "https://huggingface.co/datasets/gorilla-llm/Berkeley-Function-Calling-Leaderboard",
     "kind": "dataset",
     "note": "BFCL dataset"
    },
    {
     "name": "APIBench",
     "url": "https://huggingface.co/datasets/gorilla-llm/APIBench",
     "kind": "dataset",
     "note": "API benchmark"
    },
    {
     "name": "gorilla-openfunctions-v2",
     "url": "https://huggingface.co/gorilla-llm/gorilla-openfunctions-v2",
     "kind": "model",
     "note": "OpenFunctions v2"
    }
   ],
   "papers": [
    {
     "title": "Gorilla: Large Language Model Connected with Massive APIs",
     "url": "https://arxiv.org/abs/2305.15334",
     "year": "2023"
    },
    {
     "title": "RAFT: Adapting Language Model to Domain Specific RAG",
     "url": "https://arxiv.org/abs/2403.10131",
     "year": "2024"
    },
    {
     "title": "GoEX: Perspectives and Designs Towards a Runtime for Autonomous LLM Applications",
     "url": "https://arxiv.org/abs/2404.06921",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "gorilla",
     "url": "https://github.com/ShishirPatil/gorilla",
     "note": "Training and evaluating LLMs for function calls"
    }
   ],
   "openness": "Apache-2.0 code, models and datasets.",
   "key_people": [
    {
     "name": "Shishir G. Patil",
     "role": "Lead researcher",
     "url": "https://people.eecs.berkeley.edu/~shishirpatil/",
     "note": null
    },
    {
     "name": "Tianjun Zhang",
     "role": "Researcher",
     "url": "https://tianjunz.github.io/",
     "note": null
    },
    {
     "name": "Xin Wang",
     "role": "Researcher",
     "url": "https://xinw.ai/",
     "note": null
    },
    {
     "name": "Joseph E. Gonzalez",
     "role": "Professor, UC Berkeley",
     "url": "https://people.eecs.berkeley.edu/~jegonzal/",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the website.",
   "contact_note": "Via website.",
   "sources": [
    "https://gorilla.cs.berkeley.edu",
    "https://github.com/ShishirPatil/gorilla",
    "https://huggingface.co/gorilla-llm"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "gorilla",
   "mlat": 37.91673,
   "mlon": -122.23138,
   "logo": true
  },
  {
   "orig_id": "gpt-nl",
   "name": "GPT-NL",
   "alias": null,
   "full_name": "GPT-NL",
   "type": "nonprofit",
   "founded": null,
   "country": "Netherlands",
   "city": "The Hague",
   "lat": 52.0705,
   "lon": 4.3007,
   "website": "https://gpt-nl.nl",
   "huggingface": "https://huggingface.co/GPT-NL",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Dutch non-profit collaboration between TNO, NFI and SURF building a responsible Dutch language model, publishing its permissioned public training corpus on Hugging Face.",
   "tagline": "GPT-NL is a Dutch non-profit effort by TNO, NFI and SURF to build a lawful, sovereign Dutch language model trained on licensed data.",
   "about": [
    "GPT-NL is a joint initiative of TNO, the Netherlands Forensic Institute and SURF, launched in November 2023 with a reported budget of 13.5 million euros. It aims to supply a language model for business and public use in which sovereignty, compliance and control matter.",
    "The project stresses a clean data supply chain and has signed licensing arrangements with Dutch news publishers. It publishes datasets on Hugging Face, while its model is currently offered to launching customers, with GPT-NL version 1 scheduled for autumn 2026."
   ],
   "mission": "To provide a Dutch-language model developed responsibly, with transparent and lawfully sourced training data, for use in the Netherlands.",
   "notice": "GPT-NL's model weights are currently restricted to launch customers. The datasets it has published are open.",
   "known_for": [
    "Dutch sovereign language model",
    "Permissioned and licensed training data",
    "TNO, NFI and SURF collaboration"
   ],
   "focus": [
    "Dutch LLM",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Project launched",
     "note": "TNO, SURF and NFI started GPT-NL after a project plan awarded at the end of November 2023.",
     "year": "2023"
    },
    {
     "title": "Dutch news publishers contribute",
     "note": "Dutch news publishers began collaborating with TNO on training data through a licensing agreement.",
     "year": "2025"
    },
    {
     "title": "First pilots",
     "note": "Five organizations began feasibility studies with the beta version of GPT-NL.",
     "year": "2026"
    },
    {
     "title": "GPT-NL version 1 planned",
     "note": "The project lists the first full release for autumn 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "GPT-NL/GPT-NL_Public_Corpus",
     "url": "https://huggingface.co/datasets/GPT-NL/GPT-NL_Public_Corpus",
     "kind": "dataset",
     "note": "Dataset with 23 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "GPT-NL Public Corpus",
     "url": "https://huggingface.co/datasets/GPT-NL/GPT-NL_Public_Corpus",
     "note": "Public training corpus (CC-BY-4.0 card)"
    },
    {
     "name": "DuidelijkeTaal",
     "url": "https://huggingface.co/datasets/GPT-NL/DuidelijkeTaal-v1.0-split",
     "note": "Plain-language Dutch dataset"
    },
    {
     "name": "dutch-proverbs",
     "url": "https://huggingface.co/datasets/GPT-NL/dutch-proverbs",
     "note": "Dutch proverbs dataset"
    },
    {
     "name": "Collection metadata",
     "url": "https://huggingface.co/datasets/GPT-NL/Collection-metadata",
     "note": "Metadata for the collected corpus"
    }
   ],
   "openness": "Public corpus and datasets are open on Hugging Face; open model weights could not be verified at the time of checking.",
   "key_people": [
    {
     "name": "TNO team members",
     "role": "Contributors listed in the GPT-NL Hugging Face organization",
     "url": "https://huggingface.co/GPT-NL",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://gpt-nl.nl",
    "https://huggingface.co/GPT-NL"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "gpt-nl",
   "logo": true
  },
  {
   "orig_id": "gronlp",
   "name": "GroNLP",
   "alias": "University of Groningen",
   "full_name": "GroNLP (University of Groningen)",
   "type": "academic",
   "founded": null,
   "country": "Netherlands",
   "city": "Groningen",
   "lat": 53.2194,
   "lon": 6.5665,
   "website": "https://www.rug.nl/research/clcg/research/cl/",
   "huggingface": "https://huggingface.co/GroNLP",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computational linguistics group at the University of Groningen, known for BERTje, HateBERT and Dutch GPT-2 models.",
   "tagline": "GroNLP is the computational linguistics group at the University of Groningen, known for BERTje, HateBERT and Dutch GPT-2 models.",
   "about": [
    "GroNLP is the Computational Linguistics group of the Center for Language and Cognition Groningen. Its members, including Malvina Nissim and Arianna Bisazza, work on topics from NLP to dialectology and publish models and datasets on Hugging Face.",
    "In December 2019 the group introduced BERTje, a Dutch BERT model by Wietse de Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord and Malvina Nissim. HateBERT, a BERT retrained for abusive language detection in English, followed in October 2020."
   ],
   "mission": "To advance computational linguistics and give researchers open language models for Dutch and other languages.",
   "notice": null,
   "known_for": [
    "BERTje Dutch BERT",
    "HateBERT for abusive language",
    "Dutch GPT-2 models"
   ],
   "focus": [
    "Dutch NLP",
    "language models"
   ],
   "contributions": [
    {
     "title": "BERTje",
     "note": "A Dutch BERT model was introduced on arXiv.",
     "year": "2019"
    },
    {
     "title": "HateBERT",
     "note": "Caselli, Basile, Mitrovic and Granitzer retrained BERT for abusive language detection.",
     "year": "2020"
    },
    {
     "title": "Dutch models on Hugging Face hub",
     "note": "The GroNLP organization published bert-base-dutch-cased and gpt2-small-dutch on the hub.",
     "year": "2022"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "GroNLP/hateBERT",
     "url": "https://huggingface.co/GroNLP/hateBERT",
     "kind": "model",
     "note": "Model with 43 likes on the Hub."
    },
    {
     "name": "GroNLP/bert-base-dutch-cased",
     "url": "https://huggingface.co/GroNLP/bert-base-dutch-cased",
     "kind": "model",
     "note": "Model with 35 likes on the Hub."
    },
    {
     "name": "GroNLP/ik-nlp-22_winemag",
     "url": "https://huggingface.co/datasets/GroNLP/ik-nlp-22_winemag",
     "kind": "dataset",
     "note": "Dataset with 6 likes on the Hub."
    },
    {
     "name": "GroNLP/neural-acoustic-distance",
     "url": "https://huggingface.co/spaces/GroNLP/neural-acoustic-distance",
     "kind": "space",
     "note": "Space with 15 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "BERTje: A Dutch BERT Model",
     "url": "https://arxiv.org/abs/1912.09582",
     "year": "2019"
    },
    {
     "title": "HateBERT: Retraining BERT for Abusive Language Detection in English",
     "url": "https://arxiv.org/abs/2010.12472",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "HateBERT",
     "url": "https://huggingface.co/GroNLP/hateBERT",
     "note": "BERT retrained for abusive language detection"
    },
    {
     "name": "bert-base-dutch-cased",
     "url": "https://huggingface.co/GroNLP/bert-base-dutch-cased",
     "note": "BERTje Dutch BERT"
    },
    {
     "name": "gpt2-small-dutch",
     "url": "https://huggingface.co/GroNLP/gpt2-small-dutch",
     "note": "Dutch GPT-2"
    },
    {
     "name": "dutch-cola",
     "url": "https://huggingface.co/datasets/GroNLP/dutch-cola",
     "note": "Dutch acceptability dataset"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face.",
   "key_people": [
    {
     "name": "Malvina Nissim",
     "role": "Professor, group member",
     "url": "https://www.rug.nl/research/clcg/research/cl/?lang=en",
     "note": null
    },
    {
     "name": "Arianna Bisazza",
     "role": "Group member",
     "url": "https://www.rug.nl/research/clcg/research/cl/?lang=en",
     "note": null
    },
    {
     "name": "Andreas van Cranenburgh",
     "role": "BERTje co-author",
     "url": "https://huggingface.co/andreasvc",
     "note": null
    },
    {
     "name": "Tommaso Caselli",
     "role": "HateBERT lead author",
     "url": "https://arxiv.org/abs/2010.12472",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the university website.",
   "sources": [
    "https://www.rug.nl/research/clcg/research/cl/",
    "https://huggingface.co/GroNLP",
    "https://arxiv.org/abs/1912.09582"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "gronlp",
   "logo": true
  },
  {
   "orig_id": "h-company",
   "name": "H Company",
   "alias": null,
   "full_name": "H Company",
   "type": "startup",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.hcompany.ai",
   "huggingface": "https://huggingface.co/Hcompany",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Paris-based AI company building computer-use and web agents. It releases its Holo vision-language models as open weights on Hugging Face.",
   "tagline": "A Paris AI company building computer-use and web agents and releasing its Holo vision-language models as open weights.",
   "about": [
    "H Company was founded in Paris in 2023 by five DeepMind researchers, and Gautier Cloix became CEO in 2025. It reports raising more than 200 million euros in 2024 from backers including Accel, UiPath, Bpifrance, Bernard Arnault and Eric Schmidt.",
    "In 2025 it launched Holo1, described as the first open-weight vision-language model for computer use, together with the Surfer-H web agent. Holo1.5 and Holo2 followed in 2025 and Holo3 and Holo4 in 2026."
   ],
   "mission": "H Company aims to build agents that operate software and the web on a user's behalf, and it shares its Holo models as open weights.",
   "notice": null,
   "known_for": [
    "Holo open-weight computer-use models",
    "Surfer-H web agent",
    "Paris-based agent research"
   ],
   "focus": [
    "agents",
    "computer use",
    "vision-language models"
   ],
   "contributions": [
    {
     "title": "H Company founded",
     "note": "Five DeepMind researchers founded the company in Paris.",
     "year": "2023"
    },
    {
     "title": "Funding round",
     "note": "The company raised more than 200 million euros.",
     "year": "2024"
    },
    {
     "title": "Holo1 and Surfer-H",
     "note": "The first open-weight computer-use VLM and a web agent launched together.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "Hcompany/Holo-3.1-35B-A3B-GGUF",
     "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF",
     "kind": "model",
     "note": "Model: 64,708 downloads, 54 likes."
    },
    {
     "name": "Hcompany/Holo2-4B",
     "url": "https://huggingface.co/Hcompany/Holo2-4B",
     "kind": "model",
     "note": "Model: 25,357 downloads, 27 likes."
    },
    {
     "name": "Hcompany/trajectories",
     "url": "https://huggingface.co/datasets/Hcompany/trajectories",
     "kind": "dataset",
     "note": "Dataset: 9,356 downloads, 41 likes."
    },
    {
     "name": "Hcompany/Holo1-Navigation",
     "url": "https://huggingface.co/spaces/Hcompany/Holo1-Navigation",
     "kind": "space",
     "note": "Space: 56 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Surfer-H Meets Holo1: Cost-Efficient Web Agent Powered by Open Weights",
     "url": "https://arxiv.org/abs/2506.02865",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Holo3-35B-A3B",
     "url": "https://huggingface.co/Hcompany/Holo3-35B-A3B",
     "note": "Open-weight GUI-grounding model"
    },
    {
     "name": "Holo1-7B",
     "url": "https://huggingface.co/Hcompany/Holo1-7B",
     "note": "Open-weight 7B web navigation model"
    },
    {
     "name": "WebClick",
     "url": "https://huggingface.co/datasets/Hcompany/WebClick",
     "note": "Dataset for localisation on web pages"
    }
   ],
   "openness": "Holo models checked on Hugging Face carry Apache-2.0 licences; datasets and demo spaces are also public.",
   "key_people": [
    {
     "name": "Gautier Cloix",
     "role": "CEO",
     "url": "https://www.hcompany.ai/team",
     "note": null
    },
    {
     "name": "Géraldine Lafontaine",
     "role": "COO",
     "url": "https://www.hcompany.ai/team",
     "note": null
    },
    {
     "name": "Laurent Sifre",
     "role": "Chair of the Scientific Committee",
     "url": "https://www.hcompany.ai/team",
     "note": null
    },
    {
     "name": "Maxime Defauw",
     "role": "Head of Forward Deployed",
     "url": "https://www.hcompany.ai/team",
     "note": null
    },
    {
     "name": "Caroline Charra",
     "role": "Head of Product",
     "url": "https://www.hcompany.ai/team",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "No contact address found on the homepage; Hugging Face org is the public channel.",
   "sources": [
    "https://www.hcompany.ai",
    "https://huggingface.co/Hcompany",
    "https://arxiv.org/abs/2506.02865"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "h-company",
   "mlat": 48.84709,
   "mlon": 2.5082,
   "logo": true
  },
  {
   "orig_id": "hazy-research",
   "name": "Hazy Research",
   "alias": null,
   "full_name": "Hazy Research",
   "type": "academic",
   "founded": null,
   "country": "United States",
   "city": "Stanford",
   "lat": 37.4275,
   "lon": -122.1697,
   "website": "https://hazyresearch.stanford.edu",
   "huggingface": "https://huggingface.co/hazyresearch",
   "github": "https://github.com/HazyResearch",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Stanford group releasing open efficient-architecture research such as ThunderKittens kernels, Based and Hyena models.",
   "tagline": "A Stanford research group led by Chris Re, known for efficient architectures and GPU kernels such as ThunderKittens, Hyena, Based and Monarch Mixer.",
   "about": [
    "Hazy Research is a Stanford computer science group led by Professor Chris Re. It studies learning from weak supervision and the mathematical foundations of these methods, and turns the results into systems that help people build and maintain machine learning models.",
    "Its blog covers GPU kernel programming with ThunderKittens, HipKittens and ParallelKittens, efficient architectures including Monarch Mixer, Based, Hyena, S4 and FlashFFTConv, local on-device models such as Minions and OpenJarvis, and long-context memory techniques such as Cartridges."
   ],
   "mission": "The group aims to make machine learning systems faster and easier to build by pairing theory with open implementations.",
   "notice": null,
   "known_for": [
    "ThunderKittens GPU kernels",
    "Hyena, Based and Monarch Mixer architectures",
    "Minions local-model work"
   ],
   "focus": [
    "architectures",
    "kernels",
    "efficiency"
   ],
   "contributions": [],
   "github_highlights": [
    {
     "name": "ThunderKittens",
     "url": "https://github.com/HazyResearch/ThunderKittens",
     "note": "Tile primitives for speedy kernels."
    },
    {
     "name": "deepdive",
     "url": "https://github.com/HazyResearch/deepdive",
     "note": "Knowledge base construction system."
    },
    {
     "name": "minions",
     "url": "https://github.com/HazyResearch/minions",
     "note": "Big and small LLMs working together."
    },
    {
     "name": "data-centric-ai",
     "url": "https://github.com/HazyResearch/data-centric-ai",
     "note": "Data-centric AI resources."
    }
   ],
   "hf_highlights": [
    {
     "name": "based-1b",
     "url": "https://huggingface.co/hazyresearch/based-1b",
     "kind": "model",
     "note": "Based 1B model."
    },
    {
     "name": "based-swde",
     "url": "https://huggingface.co/datasets/hazyresearch/based-swde",
     "kind": "dataset",
     "note": "SWDE for Based evaluation."
    },
    {
     "name": "based-fda",
     "url": "https://huggingface.co/datasets/hazyresearch/based-fda",
     "kind": "dataset",
     "note": "FDA for Based evaluation."
    }
   ],
   "papers": [
    {
     "title": "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness",
     "url": "https://arxiv.org/abs/2205.14135",
     "year": "2022"
    },
    {
     "title": "Hyena Hierarchy: Towards Larger Convolutional Language Models",
     "url": "https://arxiv.org/abs/2302.10866",
     "year": "2023"
    },
    {
     "title": "Simple linear attention language models balance the recall-throughput tradeoff",
     "url": "https://arxiv.org/abs/2402.18668",
     "year": "2024"
    },
    {
     "title": "ThunderKittens: Simple, Fast, and Adorable AI Kernels",
     "url": "https://arxiv.org/abs/2410.20399",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "ThunderKittens",
     "url": "https://github.com/HazyResearch/ThunderKittens",
     "note": "GPU kernels"
    },
    {
     "name": "minions",
     "url": "https://github.com/HazyResearch/minions",
     "note": "Local+cloud LLMs"
    }
   ],
   "openness": "Open code and weights.",
   "key_people": [
    {
     "name": "Chris Ré",
     "role": "Professor and group lead",
     "url": "https://cs.stanford.edu/~chrismre/",
     "note": null
    },
    {
     "name": "Kelly Buchanan",
     "role": "Postdoctoral researcher",
     "url": "https://www.ekbuchanan.com/",
     "note": null
    },
    {
     "name": "Jon Saad-Falcon",
     "role": "PhD student",
     "url": "https://jonsaadfalcon.com/",
     "note": null
    },
    {
     "name": "Divya Nori",
     "role": "PhD student",
     "url": "https://divnori.github.io/",
     "note": null
    },
    {
     "name": "Roberto Garcia",
     "role": "PhD student",
     "url": "https://roberto09.github.io/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://hazyresearch.stanford.edu); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://hazyresearch.stanford.edu",
    "https://github.com/HazyResearch"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "hazy-research",
   "mlat": 37.3775,
   "mlon": -122.1697,
   "logo": true
  },
  {
   "orig_id": "heidelberg-nlp",
   "name": "Heidelberg NLP",
   "alias": null,
   "full_name": "Heidelberg NLP",
   "type": "academic",
   "founded": null,
   "country": "Germany",
   "city": "Heidelberg",
   "lat": 49.3988,
   "lon": 8.6724,
   "website": "https://www.cl.uni-heidelberg.de",
   "huggingface": null,
   "github": "https://github.com/Heidelberg-NLP",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computational linguistics group at Heidelberg University, releasing benchmarks and models such as VALSE and Graph Language Models.",
   "tagline": "Heidelberg NLP is the computational linguistics group at Heidelberg University, producing benchmarks such as VALSE and graph-based language models.",
   "about": [
    "The Heidelberg Natural Language Processing Group is based at the Department of Computational Linguistics, where Anette Frank is among its principal researchers. Its GitHub organization, opened in 2019, hosts code for knowledge-graph extraction, multimodal explainability and model benchmarks.",
    "Its releases include VALSE, a benchmark for vision and language models centered on linguistic phenomena, MM-SHAP for measuring multimodal contributions, and Graph Language Models, which Moritz Plenz and Anette Frank introduced in January 2024."
   ],
   "mission": "To advance language understanding and make models more faithful and explainable through open benchmarks and methods.",
   "notice": null,
   "known_for": [
    "VALSE vision-language benchmark",
    "Graph Language Models",
    "MM-SHAP and CC-SHAP explainability"
   ],
   "focus": [
    "NLP"
   ],
   "contributions": [
    {
     "title": "CoCo-Ex",
     "note": "A tool mapping text to ConceptNet concept nodes was published.",
     "year": "2021"
    },
    {
     "title": "VALSE benchmark",
     "note": "Parcalabescu, Cafagna, Muradjan, Frank, Calixto and Gatt released a task-independent vision-language benchmark.",
     "year": "2021"
    },
    {
     "title": "MM-SHAP",
     "note": "A performance-agnostic metric for multimodal contributions was released.",
     "year": "2022"
    },
    {
     "title": "CC-SHAP and ancient language models",
     "note": "Code for measuring explanation faithfulness and models for ancient languages were released.",
     "year": "2023"
    },
    {
     "title": "Graph Language Models",
     "note": "Plenz and Frank introduced language models operating on graphs.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "GraphLanguageModels",
     "url": "https://github.com/Heidelberg-NLP/GraphLanguageModels",
     "note": "Code for our paper \"Graph Language Models\" (78 stars)"
    },
    {
     "name": "CoCo-Ex",
     "url": "https://github.com/Heidelberg-NLP/CoCo-Ex",
     "note": "CoCo-Ex extracts meaningful concepts from natural language texts and maps them to conjunct concept nodes in Co (59 stars)"
    },
    {
     "name": "VALSE",
     "url": "https://github.com/Heidelberg-NLP/VALSE",
     "note": "Data repository for the VALSE benchmark. (40 stars)"
    },
    {
     "name": "ancient-language-models",
     "url": "https://github.com/Heidelberg-NLP/ancient-language-models",
     "note": "Open-source repository. (36 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "VALSE: A Task-Independent Benchmark for Vision and Language Models Centered on Linguistic Phenomena",
     "url": "https://arxiv.org/abs/2112.07566",
     "year": "2021"
    },
    {
     "title": "Graph Language Models",
     "url": "https://arxiv.org/abs/2401.07105",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "GraphLanguageModels",
     "url": "https://github.com/Heidelberg-NLP/GraphLanguageModels",
     "note": "Code for Graph Language Models"
    },
    {
     "name": "CoCo-Ex",
     "url": "https://github.com/Heidelberg-NLP/CoCo-Ex",
     "note": "Concept extraction and mapping from text"
    },
    {
     "name": "VALSE",
     "url": "https://github.com/Heidelberg-NLP/VALSE",
     "note": "Benchmark for vision-language models centred on linguistic phenomena"
    },
    {
     "name": "ancient-language-models",
     "url": "https://github.com/Heidelberg-NLP/ancient-language-models",
     "note": "Code for ancient language models"
    }
   ],
   "openness": "Code and benchmarks are public on GitHub.",
   "key_people": [
    {
     "name": "Anette Frank",
     "role": "Professor and principal researcher",
     "url": "https://github.com/AnetteFrank",
     "note": null
    },
    {
     "name": "Letitia Parcalabescu",
     "role": "VALSE first author",
     "url": "https://github.com/LetiP",
     "note": null
    },
    {
     "name": "Moritz Plenz",
     "role": "Graph Language Models first author",
     "url": "https://github.com/MoritzPlenz",
     "note": null
    },
    {
     "name": "Maria Becker",
     "role": "Researcher",
     "url": "https://github.com/maria-becker",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.cl.uni-heidelberg.de",
    "https://github.com/Heidelberg-NLP",
    "https://arxiv.org/abs/2112.07566"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "heidelberg-nlp",
   "mlat": 49.43416,
   "mlon": 8.6724,
   "logo": true
  },
  {
   "orig_id": "helsinki-nlp",
   "name": "Helsinki-NLP",
   "alias": "University of Helsinki",
   "full_name": "Helsinki-NLP (University of Helsinki)",
   "type": "academic",
   "founded": null,
   "country": "Finland",
   "city": "Helsinki",
   "lat": 60.1699,
   "lon": 24.9384,
   "website": "https://blogs.helsinki.fi/language-technology/",
   "huggingface": "https://huggingface.co/Helsinki-NLP",
   "github": "https://github.com/Helsinki-NLP",
   "x": "https://x.com/HelsinkiNLP",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Language technology research group at the University of Helsinki, creators of the OPUS-MT open translation models and OPUS parallel corpora.",
   "tagline": "The language technology group at the University of Helsinki, creators of the OPUS-MT open translation models and OPUS parallel corpora.",
   "about": [
    "Helsinki-NLP is the language technology research group at the University of Helsinki. Its research areas include cross-lingual NLP, data-intensive and neural NLP, NLP for morphologically rich languages and language technology in the humanities.",
    "The group maintains OPUS, a collection of parallel corpora, and publishes translation models and code through its GitHub and Hugging Face organisations. It takes part in national and international research projects."
   ],
   "mission": "The group aims to provide open translation models and parallel data that cover many languages, including low-resource ones.",
   "notice": null,
   "known_for": [
    "OPUS-MT open translation models",
    "OPUS parallel corpora"
   ],
   "focus": [
    "machine translation",
    "datasets",
    "multilingual"
   ],
   "contributions": [
    {
     "title": "Language technology begins at Helsinki",
     "note": "The discipline began at the Department of General Linguistics.",
     "year": "1987"
    },
    {
     "title": "Move to Modern Languages",
     "note": "The unit moved to the Department of Modern Languages.",
     "year": "2010"
    },
    {
     "title": "Move to Digital Humanities",
     "note": "The unit moved to the Department of Digital Humanities.",
     "year": "2018"
    }
   ],
   "github_highlights": [
    {
     "name": "Opus-MT",
     "url": "https://github.com/Helsinki-NLP/Opus-MT",
     "note": "Open neural machine translation models and web services (876 stars)"
    },
    {
     "name": "Tatoeba-Challenge",
     "url": "https://github.com/Helsinki-NLP/Tatoeba-Challenge",
     "note": "Tatoeba translation challenge data and models (858 stars)"
    },
    {
     "name": "OPUS-MT-train",
     "url": "https://github.com/Helsinki-NLP/OPUS-MT-train",
     "note": "Training open neural machine translation models (404 stars)"
    },
    {
     "name": "OpusFilter",
     "url": "https://github.com/Helsinki-NLP/OpusFilter",
     "note": "Parallel corpus processing toolkit (116 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "Helsinki-NLP/opus-mt-en-ru",
     "url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-ru",
     "kind": "model",
     "note": "English to Russian OPUS-MT model."
    },
    {
     "name": "Helsinki-NLP/opus-mt-zh-en",
     "url": "https://huggingface.co/Helsinki-NLP/opus-mt-zh-en",
     "kind": "model",
     "note": "Chinese to English OPUS-MT model."
    },
    {
     "name": "Helsinki-NLP/fineweb-edu-translated",
     "url": "https://huggingface.co/datasets/Helsinki-NLP/fineweb-edu-translated",
     "kind": "dataset",
     "note": "Translated FineWeb-Edu data."
    },
    {
     "name": "Helsinki-NLP/opus-translate",
     "url": "https://huggingface.co/spaces/Helsinki-NLP/opus-translate",
     "kind": "space",
     "note": "Demo space for OPUS-MT translation."
    }
   ],
   "papers": [
    {
     "title": "OPUS-MT – Building open translation services for the World",
     "url": "https://aclanthology.org/2020.eamt-1.61/",
     "year": "2020"
    },
    {
     "title": "The Tatoeba Translation Challenge -- Realistic Data Sets for Low Resource and Multilingual MT",
     "url": "https://arxiv.org/abs/2010.06354",
     "year": "2020"
    },
    {
     "title": "Democratizing Neural Machine Translation with OPUS-MT",
     "url": "https://arxiv.org/abs/2212.01936",
     "year": "2022"
    },
    {
     "title": "OpusFilter: A Configurable Parallel Corpus Filtering Toolbox",
     "url": "https://aclanthology.org/2020.acl-demos.20/",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "OPUS-MT",
     "url": "https://github.com/Helsinki-NLP/Opus-MT",
     "note": "Open neural machine translation models and services"
    },
    {
     "name": "opus-mt-zh-en",
     "url": "https://huggingface.co/Helsinki-NLP/opus-mt-zh-en",
     "note": "Chinese to English translation model"
    },
    {
     "name": "opus-100",
     "url": "https://huggingface.co/datasets/Helsinki-NLP/opus-100",
     "note": "Multilingual parallel corpus"
    },
    {
     "name": "Tatoeba Challenge",
     "url": "https://github.com/Helsinki-NLP/Tatoeba-Challenge",
     "note": "Translation benchmark data"
    }
   ],
   "openness": "Over a thousand translation models and corpora are public on Hugging Face; code is mostly MIT.",
   "key_people": [
    {
     "name": "Jörg Tiedemann",
     "role": "Group leader",
     "url": "https://blogs.helsinki.fi/language-technology/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The language technology group's blog gives group and programme information.",
   "sources": [
    "https://blogs.helsinki.fi/language-technology/",
    "https://huggingface.co/Helsinki-NLP",
    "https://github.com/Helsinki-NLP"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "helsinki-nlp",
   "logo": true
  },
  {
   "orig_id": "hitz",
   "name": "HiTZ Basque Center for Language Technology",
   "alias": null,
   "full_name": "HiTZ Basque Center for Language Technology",
   "type": "academic",
   "founded": null,
   "country": "Spain",
   "city": "Donostia-San Sebastian",
   "lat": 43.3183,
   "lon": -1.9812,
   "website": "https://www.hitz.eus",
   "huggingface": "https://huggingface.co/HiTZ",
   "github": "https://github.com/hitz-zentroa",
   "x": "https://x.com/hitz_zentroa",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Basque Center for Language Technology at the University of the Basque Country, which publishes Basque and multilingual language models and information-extraction tools.",
   "tagline": "The Basque Center for Language Technology at the University of the Basque Country, publishing Basque and multilingual language models and NLP tools.",
   "about": [
    "HiTZ is the Basque Center for Language Technology at the University of the Basque Country. It develops language technology for Basque and other languages, and its Hugging Face and GitHub organisations publish models and tools.",
    "Its releases include Basque and multilingual language models and information-extraction tools. Detailed history and staff could not be verified because the center's site blocked automated access."
   ],
   "mission": "HiTZ works to give Basque and other under-resourced languages modern language technology, openly shared.",
   "notice": null,
   "known_for": [
    "Basque language models",
    "Multilingual NLP research",
    "Information-extraction tools"
   ],
   "focus": [
    "Basque",
    "language models",
    "information extraction",
    "speech"
   ],
   "contributions": [
    {
     "title": "HiTZ research center presented",
     "note": "The Basque Government presented HiTZ as a research center grouping the Ixa and Aholab groups of the University of the Basque Country.",
     "year": "2020"
    },
    {
     "title": "Basque Government funding",
     "note": "The Basque Government announced 1.68 million euros for the center's language-technology research.",
     "year": "2023"
    },
    {
     "title": "Latxa",
     "note": "Released a family of 7B to 70B Basque language models built on Llama 2, with a new Basque corpus and evaluation suite, under the IKER-GAITU project.",
     "year": "2024"
    },
    {
     "title": "Latxa Instruct on Llama 3.1",
     "note": "Published instruction-tuned Latxa models for Basque at 8B and 70B and a public chatbot, with a crowdsourced arena collecting over 12,000 preference judgments.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "GoLLIE",
     "url": "https://github.com/hitz-zentroa/GoLLIE",
     "note": "Guideline-following LLM for information extraction (445 stars)"
    },
    {
     "name": "lm-contamination",
     "url": "https://github.com/hitz-zentroa/lm-contamination",
     "note": "Database of benchmark contamination evidence (82 stars)"
    },
    {
     "name": "whisper-lm",
     "url": "https://github.com/hitz-zentroa/whisper-lm",
     "note": "N-gram and LLM support for Whisper models (44 stars)"
    },
    {
     "name": "latxa",
     "url": "https://github.com/hitz-zentroa/latxa",
     "note": "Open language model and evaluation suite for Basque (38 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "HiTZ/Latxa-Llama-3.1-8B-Instruct",
     "url": "https://huggingface.co/HiTZ/Latxa-Llama-3.1-8B-Instruct",
     "kind": "model",
     "note": "Basque instruct model on Llama 3.1."
    },
    {
     "name": "HiTZ/Latxa-Llama-3.1-70B-Instruct-v2-FP8",
     "url": "https://huggingface.co/HiTZ/Latxa-Llama-3.1-70B-Instruct-v2-FP8",
     "kind": "model",
     "note": "70B Basque instruct model in FP8."
    },
    {
     "name": "HiTZ/MedExpQA",
     "url": "https://huggingface.co/datasets/HiTZ/MedExpQA",
     "kind": "dataset",
     "note": "Medical multilingual question answering benchmark."
    },
    {
     "name": "HiTZ/EusExams",
     "url": "https://huggingface.co/datasets/HiTZ/EusExams",
     "kind": "dataset",
     "note": "Basque exam questions for evaluation."
    }
   ],
   "papers": [
    {
     "title": "Latxa: An Open Language Model and Evaluation Suite for Basque",
     "url": "https://arxiv.org/abs/2403.20266",
     "year": "2024"
    },
    {
     "title": "GoLLIE: Annotation Guidelines improve Zero-Shot Information-Extraction",
     "url": "https://arxiv.org/abs/2310.03668",
     "year": "2023"
    },
    {
     "title": "NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark",
     "url": "https://arxiv.org/abs/2310.18018",
     "year": "2023"
    },
    {
     "title": "This is not a Dataset: A Large Negation Benchmark to Challenge Large Language Models",
     "url": "https://arxiv.org/abs/2310.15941",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "GoLLIE-7B",
     "url": "https://huggingface.co/HiTZ/GoLLIE-7B",
     "note": "Guideline-following information extraction LLM"
    },
    {
     "name": "Latxa",
     "url": "https://github.com/hitz-zentroa/latxa",
     "note": "Open language model and evaluation suite for Basque"
    },
    {
     "name": "Multilingual Medical Corpus",
     "url": "https://huggingface.co/datasets/HiTZ/Multilingual-Medical-Corpus",
     "note": "Medical corpus"
    }
   ],
   "openness": "Models, datasets and code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Eneko Agirre",
     "role": "Director of HiTZ; Professor, University of the Basque Country",
     "url": "https://eagirre.github.io/",
     "note": "Principal investigator of the Ixa group and co-author of the Latxa paper."
    },
    {
     "name": "German Rigau",
     "role": "Deputy Director of HiTZ",
     "url": "https://www.hitz.eus/sites/default/files/HiTZ_2025_report.pdf",
     "note": "Named as deputy director in the center's annual report; co-author of the Latxa paper."
    },
    {
     "name": "Julen Etxaniz",
     "role": "First author of Latxa",
     "url": "https://arxiv.org/abs/2403.20266",
     "note": "Led the Basque model and evaluation suite paper."
    }
   ],
   "discord_note": null,
   "contact_note": "The centre's website could not be fetched during checking; use its GitHub or Hugging Face organisation for project contact.",
   "sources": [
    "https://www.hitz.eus",
    "https://huggingface.co/HiTZ",
    "https://github.com/hitz-zentroa",
    "https://arxiv.org/abs/2403.20266"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "hitz-basque-center-for-language-technology",
   "logo": true
  },
  {
   "orig_id": "hooshvare",
   "name": "Hooshvare Research Lab",
   "alias": null,
   "full_name": "Hooshvare Research Lab",
   "type": "nonprofit",
   "founded": null,
   "country": "Iran",
   "city": "Tehran",
   "lat": 35.69,
   "lon": 51.39,
   "website": "https://github.com/hooshvare/",
   "huggingface": "https://huggingface.co/HooshvareLab",
   "github": "https://github.com/hooshvare",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Non-profit lab publishing Persian NLP models such as ParsBERT and Persian GPT-2.",
   "tagline": "A small non-profit Persian NLP lab that published ParsBERT, the first widely used monolingual BERT for Persian, along with Persian GPT-2 and NER models.",
   "about": [
    "Hooshvare Research Lab is a Tehran-based group that publishes open models and datasets for Persian natural language processing. Its GitHub organization describes it as a platform for sharing information about artificial intelligence concepts with enthusiasts. The lab's work is published through the Hooshvare organizations on GitHub and Hugging Face.",
    "Its best-known release is ParsBERT, a monolingual BERT-style model for Persian described in a May 2020 preprint by Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani and Mohammad Manthouri. The paper reports gains over multilingual BERT on sentiment analysis, text classification and named entity recognition. The lab later added a Persian GPT-2 model, NER models, a news summarization dataset and workshop materials.",
    "Code and weights are released under permissive licenses such as MIT and Apache 2.0, which has made the models a common baseline for Persian-language research and applications."
   ],
   "mission": "The lab aims to make capable language models and datasets for Persian freely available, so that researchers and developers working in the language are not limited to multilingual models built for other languages.",
   "notice": null,
   "known_for": [
    "ParsBERT, a monolingual Persian BERT",
    "Persian GPT-2 and NER models",
    "The pn-summary Persian summarization dataset"
   ],
   "focus": [
    "Persian",
    "NLP"
   ],
   "contributions": [
    {
     "title": "ParsBERT preprint",
     "note": "Introduced a Persian-specific BERT-Base model that outperformed multilingual BERT on three downstream tasks.",
     "year": "2020"
    },
    {
     "title": "pn-summary dataset",
     "note": "Released a Persian news summarization dataset alongside ParsBERT-based summarization work.",
     "year": "2020"
    },
    {
     "title": "ParsBERT in Neural Processing Letters",
     "note": "Peer-reviewed journal version of the ParsBERT work published in volume 53.",
     "year": "2021"
    },
    {
     "title": "Persian GPT-2 and NER models",
     "note": "Published ParsGPT and ParsNER repositories to extend the Persian model family beyond encoders.",
     "year": "2021"
    },
    {
     "title": "ParsBERT workshop",
     "note": "Published workshop material covering ParsBERT's architecture and applications.",
     "year": "2021"
    }
   ],
   "github_highlights": [
    {
     "name": "parsbert",
     "url": "https://github.com/hooshvare/parsbert",
     "note": "Persian BERT."
    },
    {
     "name": "pn-summary",
     "url": "https://github.com/hooshvare/pn-summary",
     "note": "Persian summarization dataset."
    },
    {
     "name": "parsgpt",
     "url": "https://github.com/hooshvare/parsgpt",
     "note": "Persian GPT2."
    },
    {
     "name": "pdf2word",
     "url": "https://github.com/hooshvare/pdf2word",
     "note": "Persian PDF text extraction."
    }
   ],
   "hf_highlights": [
    {
     "name": "HooshvareLab/bert-fa-zwnj-base",
     "url": "https://huggingface.co/HooshvareLab/bert-fa-zwnj-base",
     "kind": "model",
     "note": "Persian BERT."
    },
    {
     "name": "HooshvareLab/bert-fa-base-uncased",
     "url": "https://huggingface.co/HooshvareLab/bert-fa-base-uncased",
     "kind": "model",
     "note": "Persian BERT."
    },
    {
     "name": "HooshvareLab/bert-base-parsbert-uncased",
     "url": "https://huggingface.co/HooshvareLab/bert-base-parsbert-uncased",
     "kind": "model",
     "note": "ParsBERT."
    },
    {
     "name": "HooshvareLab/pn_summary",
     "url": "https://huggingface.co/datasets/HooshvareLab/pn_summary",
     "kind": "dataset",
     "note": "Summarization."
    }
   ],
   "papers": [
    {
     "title": "ParsBERT: Transformer-based Model for Persian Language Understanding",
     "url": "https://arxiv.org/abs/2005.12515",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "gpt2-fa",
     "url": "https://huggingface.co/HooshvareLab/gpt2-fa",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "roberta-fa-zwnj-base",
     "url": "https://huggingface.co/HooshvareLab/roberta-fa-zwnj-base",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open models.",
   "key_people": [
    {
     "name": "Mehrdad Farahani",
     "role": "Lead author of ParsBERT",
     "url": "https://github.com/m3hrdadfi",
     "note": "Public member of the Hooshvare GitHub organization"
    },
    {
     "name": "Mohammad Manthouri",
     "role": "Co-author of the ParsBERT paper",
     "url": "https://github.com/mmanthouri",
     "note": "Public member of the Hooshvare GitHub organization"
    },
    {
     "name": "Mohammad Gharachorloo",
     "role": "Co-author of the ParsBERT paper",
     "url": "https://arxiv.org/abs/2005.12515",
     "note": null
    },
    {
     "name": "Marzieh Farahani",
     "role": "Co-author of the ParsBERT paper",
     "url": "https://arxiv.org/abs/2005.12515",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The GitHub org page is the lab's public home; use issues on its repos.",
   "sources": [
    "https://huggingface.co/HooshvareLab",
    "https://github.com/hooshvare/"
   ],
   "confidence": "high",
   "region": "Middle East",
   "id": "hooshvare-research-lab",
   "mlat": 35.65464,
   "mlon": 51.39,
   "logo": true
  },
  {
   "orig_id": "hpc-ai-tech",
   "name": "HPC-AI Tech",
   "alias": "Colossal-AI",
   "full_name": "HPC-AI Tech (Colossal-AI)",
   "type": "startup",
   "founded": 2021,
   "country": "Singapore",
   "city": "Singapore",
   "lat": 1.35,
   "lon": 103.82,
   "website": "https://hpcaitech.com/",
   "huggingface": "https://huggingface.co/hpcai-tech",
   "github": "https://github.com/hpcaitech",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "service@hpc-ai.com",
   "summary": "Company behind Colossal-AI distributed training tooling and the Open-Sora video generation models.",
   "tagline": "Singapore company behind Colossal-AI, an open-source stack for large-model training, and the Open-Sora video generation models.",
   "about": [
    "HPC-AI Tech was founded by Yang You, a professor at the National University of Singapore who earned his PhD at UC Berkeley. The company grew out of research on efficient distributed training and became known for Colossal-AI, an open-source software stack for training and serving large models. It raised a seed round of about six million dollars in 2022 and a 50 million dollar Series A announced in 2024.",
    "Its open-source output includes Colossal-AI and Open-Sora, a video generation project that attracted wide attention on GitHub. Alongside these, the company operates a commercial GPU cloud platform and a model platform with OpenAI-compatible and Anthropic-compatible APIs, and offers enterprise services such as fine-tuning and performance optimization.",
    "The business model combines free open-source infrastructure with paid cloud and services, so the open projects serve as the technical base of the commercial offering."
   ],
   "mission": "The company aims to lower the cost and complexity of training and running large AI models, and to keep the core tooling open while selling cloud and enterprise services on top.",
   "notice": null,
   "known_for": [
    "Colossal-AI large-model training stack",
    "Open-Sora open video generation",
    "Founded by NUS professor Yang You"
   ],
   "focus": [
    "training infrastructure",
    "video generation"
   ],
   "contributions": [
    {
     "title": "Colossal-AI",
     "note": "Open-sourced a distributed training system that unifies data, tensor and pipeline parallelism for large models.",
     "year": "2021"
    },
    {
     "title": "Seed and angel round",
     "note": "Completed a seed and angel financing round of about six million dollars led by BlueRun Ventures.",
     "year": "2022"
    },
    {
     "title": "Open-Sora",
     "note": "Released an open video generation project that reached the top of GitHub trending.",
     "year": "2024"
    },
    {
     "title": "Series A",
     "note": "Announced a 50 million dollar Series A to expand video generation and its GPU platform.",
     "year": "2024"
    },
    {
     "title": "HPC-AI.COM cloud platform",
     "note": "Offers pay-as-you-go access to datacenter GPUs and a hosted model platform.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "ColossalAI",
     "url": "https://github.com/hpcaitech/ColossalAI",
     "note": "Making large AI models cheaper, faster and more accessible."
    },
    {
     "name": "Open-Sora",
     "url": "https://github.com/hpcaitech/Open-Sora",
     "note": "Democratizing efficient video production."
    },
    {
     "name": "EnergonAI",
     "url": "https://github.com/hpcaitech/EnergonAI",
     "note": "Large-scale model inference."
    },
    {
     "name": "FastFold",
     "url": "https://github.com/hpcaitech/FastFold",
     "note": "Optimising AlphaFold training and inference."
    }
   ],
   "hf_highlights": [
    {
     "name": "Open-Sora-v2",
     "url": "https://huggingface.co/hpcai-tech/Open-Sora-v2",
     "kind": "model",
     "note": "Open-Sora 2.0 video model."
    },
    {
     "name": "OpenSora-VAE-v1.2",
     "url": "https://huggingface.co/hpcai-tech/OpenSora-VAE-v1.2",
     "kind": "model",
     "note": "Video VAE."
    },
    {
     "name": "OpenSora-STDiT-v3",
     "url": "https://huggingface.co/hpcai-tech/OpenSora-STDiT-v3",
     "kind": "model",
     "note": "Diffusion transformer."
    },
    {
     "name": "Colossal-LLaMA-2-7b-base",
     "url": "https://huggingface.co/hpcai-tech/Colossal-LLaMA-2-7b-base",
     "kind": "model",
     "note": "Continued pre-trained LLaMA-2."
    }
   ],
   "papers": [
    {
     "title": "Colossal-AI: A Unified Deep Learning System For Large-Scale Parallel Training",
     "url": "https://arxiv.org/abs/2110.14883",
     "year": "2021"
    },
    {
     "title": "Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k",
     "url": "https://arxiv.org/abs/2503.09642",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Open-Sora-v2",
     "url": "https://huggingface.co/hpcai-tech/Open-Sora-v2",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "OpenSora-VAE-v1.3",
     "url": "https://huggingface.co/hpcai-tech/OpenSora-VAE-v1.3",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open-source code and model weights.",
   "key_people": [
    {
     "name": "Yang You",
     "role": "Founder and chairman; Presidential Young Professor, National University of Singapore",
     "url": "https://www.comp.nus.edu.sg/~youy",
     "note": "UC Berkeley PhD; leads the HPC-AI Lab at NUS."
    },
    {
     "name": "James Demmel",
     "role": "Strategic adviser",
     "url": "https://company.hpc-ai.com/blog/singapore-startup-hpc-ai-tech-secures-50-million-usd-in-series-a-funding-to-build-the-video-generation-ai-model-and-gpu-platform",
     "note": "Named by the company as providing strategic guidance; Berkeley professor."
    }
   ],
   "discord_note": null,
   "contact_note": "Its site prints a service email address; code questions go through GitHub issues.",
   "sources": [
    "https://huggingface.co/hpcai-tech",
    "https://hpcaitech.com/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "hpc-ai-tech",
   "mlat": 1.38536,
   "mlon": 103.82,
   "logo": true
  },
  {
   "orig_id": "hplt",
   "name": "HPLT",
   "alias": "High Performance Language Technologies",
   "full_name": "HPLT (High Performance Language Technologies)",
   "type": "institute",
   "founded": 2022,
   "country": "Czechia",
   "city": "Prague",
   "lat": 50.0755,
   "lon": 14.4378,
   "website": "https://hplt-project.org",
   "huggingface": "https://huggingface.co/HPLT",
   "github": "https://github.com/hplt-project",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "EU-funded consortium coordinated from Charles University that publishes petabyte-scale multilingual web corpora, translation models and data-cleaning tools.",
   "tagline": "HPLT is an EU-funded consortium coordinated by Charles University that publishes petabyte-scale multilingual web corpora and models.",
   "about": [
    "High Performance Language Technologies is a three-year EU-funded project that started in September 2022. It is coordinated by Charles University in Prague and brings together five universities, two high-performance computing centers and a private NLP company.",
    "Its first release covered monolingual data in 75 languages at 5.6 trillion tokens and parallel data in 18 language pairs. The second release reached 7.6 trillion tokens in 193 languages and 380 million parallel sentences in 51 pairs, distributed under the CC0 license on Hugging Face and its website."
   ],
   "mission": "To provide open multilingual datasets, translation models and tools for training language models at scale.",
   "notice": null,
   "known_for": [
    "Petabyte-scale multilingual web corpora",
    "HPLT v2 CC0 datasets",
    "OpusCleaner and OpusTrainer tools"
   ],
   "focus": [
    "multilingual datasets",
    "machine translation",
    "language models"
   ],
   "contributions": [
    {
     "title": "Project starts",
     "note": "The EU-funded HPLT project began in September 2022 under Charles University.",
     "year": "2022"
    },
    {
     "title": "OpusCleaner",
     "note": "A web interface for cleaning machine translation training data was published.",
     "year": "2022"
    },
    {
     "title": "First data and model release",
     "note": "HPLT published version 1 datasets and models, described at EAMT 2024.",
     "year": "2024"
    },
    {
     "title": "HPLT-MT models",
     "note": "Configuration and scripts for HPLT translation model releases were published.",
     "year": "2024"
    },
    {
     "title": "HPLT v2 data release",
     "note": "The second release of cleaned monolingual and parallel data under CC0 was described at MT Summit 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "sacremoses",
     "url": "https://github.com/hplt-project/sacremoses",
     "note": "Python port of Moses tokenizer, truecaser and normalizer (498 stars)"
    },
    {
     "name": "OpusCleaner",
     "url": "https://github.com/hplt-project/OpusCleaner",
     "note": "OpusCleaner is a web interface that helps you select, clean and schedule your data for training machine transl (58 stars)"
    },
    {
     "name": "OpusTrainer",
     "url": "https://github.com/hplt-project/OpusTrainer",
     "note": "Curriculum training (25 stars)"
    },
    {
     "name": "data-analytics-tool",
     "url": "https://github.com/hplt-project/data-analytics-tool",
     "note": "HPLT Analytics (16 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "HPLT/OpenLID-v3",
     "url": "https://huggingface.co/HPLT/OpenLID-v3",
     "kind": "model",
     "note": "Model with 5 likes on the Hub."
    },
    {
     "name": "HPLT/translate-en-hr-v1.0-hplt_opus",
     "url": "https://huggingface.co/HPLT/translate-en-hr-v1.0-hplt_opus",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "HPLT/HPLT2.0_cleaned",
     "url": "https://huggingface.co/datasets/HPLT/HPLT2.0_cleaned",
     "kind": "dataset",
     "note": "Dataset with 55 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "An Expanded Massive Multilingual Dataset for High-Performance Language Technologies (HPLT)",
     "url": "https://arxiv.org/abs/2503.10267",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "HPLT 3.0",
     "url": "https://huggingface.co/datasets/HPLT/HPLT3.0",
     "note": "Multilingual web corpus on the Hub"
    },
    {
     "name": "HPLT 2.0 cleaned",
     "url": "https://huggingface.co/datasets/HPLT/HPLT2.0_cleaned",
     "note": "Cleaned version of the HPLT 2.0 corpus"
    },
    {
     "name": "OpusCleaner",
     "url": "https://github.com/hplt-project/OpusCleaner",
     "note": "Web interface for cleaning translation data"
    },
    {
     "name": "OpenLID",
     "url": "https://huggingface.co/HPLT/OpenLID-v3",
     "note": "Language identification model"
    }
   ],
   "openness": "Datasets, models and tools are public (the HPLT3.0 card lists CC0).",
   "key_people": [
    {
     "name": "Stephan Oepen",
     "role": "HPLT member (University of Oslo)",
     "url": "https://huggingface.co/oepen",
     "note": null
    },
    {
     "name": "Nikolay Bogoychev",
     "role": "HPLT member",
     "url": "https://huggingface.co/XapaJIaMnu",
     "note": null
    },
    {
     "name": "David Samuel",
     "role": "HPLT member",
     "url": "https://huggingface.co/davda54",
     "note": null
    },
    {
     "name": "Jelmer van der Linde",
     "role": "HPLT member",
     "url": "https://huggingface.co/jelmervdl",
     "note": null
    },
    {
     "name": "Pinzhen Chen",
     "role": "HPLT member",
     "url": "https://huggingface.co/pinzhenchen",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the project website.",
   "sources": [
    "https://hplt-project.org",
    "https://huggingface.co/HPLT",
    "https://github.com/hplt-project",
    "https://arxiv.org/abs/2503.10267"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "hplt",
   "mlat": 50.09715,
   "mlon": 14.49623,
   "logo": true
  },
  {
   "orig_id": "hugging-face",
   "name": "Hugging Face",
   "alias": null,
   "full_name": "Hugging Face",
   "type": "startup",
   "founded": 2016,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://huggingface.co",
   "huggingface": "https://huggingface.co/huggingface",
   "github": "https://github.com/huggingface",
   "x": "https://x.com/huggingface",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/huggingface",
   "contact_email": "press@huggingface.co",
   "summary": "The company behind the Hugging Face Hub and open-source libraries such as Transformers, Diffusers and Datasets. Headquartered in New York with a large Paris presence, it hosts the main community platform for open models and datasets.",
   "tagline": "The company behind the Hugging Face Hub and the Transformers, Diffusers and Datasets libraries, the central platform for sharing open models and datasets.",
   "about": [
    "Hugging Face was founded in 2016 in New York by Clément Delangue, Julien Chaumond and Thomas Wolf. It began as a chatbot app for teenagers and shifted to machine learning tooling after the team open-sourced the model behind the app. Delangue is chief executive, Chaumond is chief technology officer and Wolf is chief science officer.",
    "The company's open-source libraries, notably Transformers, became standard tools for working with pretrained models, and the Hub grew into the main place where researchers and companies publish models and datasets. In 2021 it launched the BigScience workshop, which produced the 176-billion-parameter multilingual BLOOM model in 2022. It also has a large presence in Paris and has expanded into robotics through the 2025 acquisition of Pollen Robotics.",
    "Hugging Face earns revenue from hosted services and enterprise offerings while keeping its libraries and much of its research open. In September 2026 Nvidia announced an agreement to acquire the company, pending regulatory approval."
   ],
   "mission": "Hugging Face states its aim as democratizing good machine learning by making models, datasets and tools open and shared by a community.",
   "notice": "In September 2026 Nvidia announced a definitive agreement to acquire Hugging Face. The transaction is expected to close in the first half of 2027, and Nvidia has said the platform will remain open and keep operating under the Hugging Face brand.",
   "known_for": [
    "The Hugging Face Hub for open models and datasets",
    "The Transformers library",
    "Coordinating the BigScience BLOOM project"
   ],
   "focus": [
    "model hub",
    "libraries",
    "datasets",
    "small language models",
    "robotics"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in New York as a chatbot app before pivoting to machine learning infrastructure.",
     "year": "2016"
    },
    {
     "title": "Transformers library",
     "note": "Released the library that unified access to pretrained language models across frameworks.",
     "year": "2019"
    },
    {
     "title": "BigScience workshop",
     "note": "Launched a year-long open research workshop to build a large multilingual language model.",
     "year": "2021"
    },
    {
     "title": "BLOOM",
     "note": "BigScience announced BLOOM, a 176-billion-parameter multilingual open model.",
     "year": "2022"
    },
    {
     "title": "Series D funding",
     "note": "Raised 235 million dollars at a 4.5 billion dollar valuation.",
     "year": "2023"
    },
    {
     "title": "Pollen Robotics acquisition",
     "note": "Acquired the French humanoid robotics startup Pollen Robotics.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "transformers",
     "url": "https://github.com/huggingface/transformers",
     "note": "Flagship model-definition library, by far the org's most-starred repo."
    },
    {
     "name": "pytorch-image-models",
     "url": "https://github.com/huggingface/pytorch-image-models",
     "note": "timm, the largest collection of PyTorch image backbones."
    },
    {
     "name": "diffusers",
     "url": "https://github.com/huggingface/diffusers",
     "note": "State-of-the-art diffusion model library."
    },
    {
     "name": "smolagents",
     "url": "https://github.com/huggingface/smolagents",
     "note": "Minimal library for agents that think in code."
    }
   ],
   "hf_highlights": [
    {
     "name": "ai-deadlines",
     "url": "https://huggingface.co/spaces/huggingface/ai-deadlines",
     "kind": "space",
     "note": "Space tracking AI conference deadlines."
    },
    {
     "name": "open-source-ai-year-in-review-2024",
     "url": "https://huggingface.co/spaces/huggingface/open-source-ai-year-in-review-2024",
     "kind": "space",
     "note": "Year-in-review Space on open source AI."
    },
    {
     "name": "documentation-images",
     "url": "https://huggingface.co/datasets/huggingface/documentation-images",
     "kind": "dataset",
     "note": "Images used across Hugging Face documentation."
    }
   ],
   "papers": [
    {
     "title": "HuggingFace's Transformers: State-of-the-art Natural Language Processing",
     "url": "https://arxiv.org/abs/1910.03771",
     "year": "2019"
    },
    {
     "title": "Datasets: A Community Library for Natural Language Processing",
     "url": "https://arxiv.org/abs/2109.02846",
     "year": "2021"
    },
    {
     "title": "SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model",
     "url": "https://arxiv.org/abs/2502.02737",
     "year": "2025"
    },
    {
     "title": "The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale",
     "url": "https://arxiv.org/abs/2406.17557",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Transformers",
     "url": "https://github.com/huggingface/transformers",
     "note": "Model-definition library for text, vision, audio and multimodal models"
    },
    {
     "name": "Diffusers",
     "url": "https://github.com/huggingface/diffusers",
     "note": "Library for diffusion models"
    },
    {
     "name": "Datasets",
     "url": "https://github.com/huggingface/datasets",
     "note": "Library and hub for ready-to-use datasets"
    },
    {
     "name": "SmolLM2 paper",
     "url": "https://arxiv.org/abs/2502.02737",
     "note": "Small language model trained with open data recipes"
    }
   ],
   "openness": "Core libraries are Apache-2.0; the Hub hosts open models and datasets and the company publishes its own models, datasets and training recipes openly.",
   "key_people": [
    {
     "name": "Clément Delangue",
     "role": "Co-founder and CEO",
     "url": "https://huggingface.co/clem",
     "note": null
    },
    {
     "name": "Julien Chaumond",
     "role": "Co-founder and CTO",
     "url": "https://huggingface.co/julien-c",
     "note": null
    },
    {
     "name": "Thomas Wolf",
     "role": "Co-founder and Chief Science Officer",
     "url": "https://huggingface.co/thomwolf",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://huggingface.co); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co",
    "https://huggingface.co/huggingface",
    "https://github.com/huggingface",
    "https://arxiv.org/abs/1910.03771"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "hugging-face",
   "mlat": 48.88481,
   "mlon": 2.50289,
   "logo": true
  },
  {
   "orig_id": "nytk",
   "name": "Hungarian Research Centre for Linguistics",
   "alias": "NYTK, PULI",
   "full_name": "Hungarian Research Centre for Linguistics (NYTK, PULI)",
   "type": "institute",
   "founded": null,
   "country": "Hungary",
   "city": "Budapest",
   "lat": 47.4979,
   "lon": 19.0402,
   "website": "https://nlp.nytud.hu",
   "huggingface": "https://huggingface.co/NYTK",
   "github": "https://github.com/nytud",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Language technology research group that publishes the PULI family of Hungarian language models along with corpora and the e-magyar text processing system.",
   "tagline": "The Hungarian Research Centre for Linguistics' language technology group publishes the PULI Hungarian language models and national corpora.",
   "about": [
    "The group grew from the Corpus Linguistics Department formed in 1997 and is led by Enikő Héja, with Gábor Prószéky and Tamás Váradi among its senior staff. It builds language resources, tools and neural models for Hungarian.",
    "Its work includes the Hungarian National Corpus, completed in 2005 with a second edition in 2014, the e-magyar text processing system, and the HuLU benchmark modeled on GLUE. Its PULI family includes PULI-GPT-3SX, a seven-billion-parameter Hungarian model."
   ],
   "mission": "To develop language technology for Hungarian, from corpora and morphology to modern neural language models.",
   "notice": null,
   "known_for": [
    "PULI Hungarian language models",
    "Hungarian National Corpus",
    "e-magyar toolchain and HuLU"
   ],
   "focus": [
    "Hungarian NLP",
    "language models",
    "corpora"
   ],
   "contributions": [
    {
     "title": "Corpus Linguistics Department formed",
     "note": "The predecessor of the group was established.",
     "year": "1997"
    },
    {
     "title": "First Hungarian National Corpus",
     "note": "The MNSZ corpus was completed.",
     "year": "2005"
    },
    {
     "title": "MNSZ2 released",
     "note": "The second Hungarian National Corpus was released.",
     "year": "2014"
    },
    {
     "title": "e-magyar toolchain",
     "note": "The emtsv text processing system repository was opened.",
     "year": "2018"
    },
    {
     "title": "HuLU benchmark and PULI-GPT-3SX",
     "note": "A Hungarian GLUE-style benchmark kit and a 7B Hungarian GPT-3 model were published.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "emtsv",
     "url": "https://github.com/nytud/emtsv",
     "note": "e-magyar text processing system -- inter-module communication via tsv + REST API (33 stars)"
    },
    {
     "name": "hadifogoly-adatbazis",
     "url": "https://github.com/nytud/hadifogoly-adatbazis",
     "note": "A magyar hadifoglyok adatbázisának orosz-magyar transzkripciója (23 stars)"
    },
    {
     "name": "emMorph",
     "url": "https://github.com/nytud/emMorph",
     "note": "Open-source repository. (18 stars)"
    },
    {
     "name": "NYTK-NerKor",
     "url": "https://github.com/nytud/NYTK-NerKor",
     "note": "The home repository of the NerKor corpus, a Hungarian gold standard named entity annotated corpus containing 1 (16 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "NYTK/named-entity-recognition-nerkor-hubert-hungarian",
     "url": "https://huggingface.co/NYTK/named-entity-recognition-nerkor-hubert-hungarian",
     "kind": "model",
     "note": "Model with 16 likes on the Hub."
    },
    {
     "name": "NYTK/PULI-GPT-3SX",
     "url": "https://huggingface.co/NYTK/PULI-GPT-3SX",
     "kind": "model",
     "note": "Model with 13 likes on the Hub."
    },
    {
     "name": "NYTK/alpaca_hu_2k",
     "url": "https://huggingface.co/datasets/NYTK/alpaca_hu_2k",
     "kind": "dataset",
     "note": "Dataset with 6 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "PULI-GPT-3SX",
     "url": "https://huggingface.co/NYTK/PULI-GPT-3SX",
     "note": "Hungarian GPT model"
    },
    {
     "name": "PULI-LlumiX-32K",
     "url": "https://huggingface.co/NYTK/PULI-LlumiX-32K",
     "note": "Hungarian-English long-context model"
    },
    {
     "name": "emtsv",
     "url": "https://github.com/nytud/emtsv",
     "note": "e-magyar text processing system"
    },
    {
     "name": "NerKor",
     "url": "https://github.com/nytud/NYTK-NerKor",
     "note": "Hungarian NER corpus"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face and GitHub; some model cards use Llama-based licences.",
   "key_people": [
    {
     "name": "Enikő Héja",
     "role": "Research group leader",
     "url": "https://nlp.nytud.hu/",
     "note": null
    },
    {
     "name": "Gábor Prószéky",
     "role": "Chief director and scientific adviser",
     "url": "https://nlp.nytud.hu/",
     "note": null
    },
    {
     "name": "Tamás Váradi",
     "role": "Director and senior research fellow",
     "url": "https://huggingface.co/VarTamas",
     "note": null
    },
    {
     "name": "Noémi Ligeti-Nagy",
     "role": "Research fellow",
     "url": "https://huggingface.co/ligetinagy",
     "note": null
    },
    {
     "name": "Zijian Győző Yang",
     "role": "Research fellow",
     "url": "https://huggingface.co/fragata",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the research group website.",
   "sources": [
    "https://nlp.nytud.hu",
    "https://huggingface.co/NYTK",
    "https://github.com/nytud"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "hungarian-research-centre-for-linguistics",
   "logo": true
  },
  {
   "orig_id": "idiap",
   "name": "Idiap Research Institute",
   "alias": null,
   "full_name": "Idiap Research Institute",
   "type": "institute",
   "founded": 1991,
   "country": "Switzerland",
   "city": "Martigny",
   "lat": 46.1029,
   "lon": 7.0721,
   "website": "https://www.idiap.ch",
   "huggingface": "https://huggingface.co/Idiap",
   "github": "https://github.com/idiap",
   "x": "https://x.com/Idiap_ch",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/idiap-research-institute",
   "contact_email": null,
   "summary": "Independent Swiss non-profit research institute in speech, vision and machine learning that publishes extensive open-source code.",
   "tagline": "Independent Swiss non-profit research institute in Martigny working on speech, vision and machine learning, and the home of the original Torch library.",
   "about": [
    "Idiap Research Institute was created in 1991 as an autonomous non-profit foundation in Martigny, Switzerland, at the initiative of the Dalle Molle Foundation. Its founding partners include the City of Martigny, the Canton of Valais, EPFL, the University of Geneva and Swisscom. The Swiss Confederation recognizes it as an institute of national importance.",
    "The institute researches speech and audio processing, computer vision, biometrics and machine learning under the tagline Artificial Intelligence for Society. In 2002 researchers there developed Torch, an early machine learning library that is a forerunner of PyTorch. Idiap also coordinated the national research center on interactive multimodal information management from 2001 to 2013, spun out companies such as Klewel in 2007, and has become a certified biometric testing center.",
    "The institute publishes a large amount of open-source code and datasets on GitHub and Hugging Face. As of 2026 it is led by an interim management team while it recruits a new director."
   ],
   "mission": "Idiap aims to advance research, innovation and education in artificial intelligence for the benefit of society, while remaining a non-profit and independent institute.",
   "notice": null,
   "known_for": [
    "Torch, an early forerunner of PyTorch",
    "Speech, vision and biometrics research",
    "Extensive open-source research software"
   ],
   "focus": [
    "speech",
    "vision",
    "machine learning"
   ],
   "contributions": [
    {
     "title": "Institute founded",
     "note": "Created in Martigny as an autonomous non-profit research foundation.",
     "year": "1991"
    },
    {
     "title": "NCCR IM2",
     "note": "Began coordinating a Swiss national research center on interactive multimodal information management, which ran until 2013.",
     "year": "2001"
    },
    {
     "title": "Torch",
     "note": "Researchers Samy Bengio, Ronan Collobert and Johnny Mariéthoz developed the Torch machine learning library.",
     "year": "2002"
    },
    {
     "title": "Klewel spin-off",
     "note": "Founded the institute's first spin-off company.",
     "year": "2007"
    },
    {
     "title": "Idiap Create Challenge",
     "note": "Hosted the first nine-day AI entrepreneurship hackathon.",
     "year": "2011"
    }
   ],
   "github_highlights": [
    {
     "name": "coqui-ai-TTS",
     "url": "https://github.com/idiap/coqui-ai-TTS",
     "note": "Fork of the Coqui deep learning TTS toolkit (2,357 stars)"
    },
    {
     "name": "fast-transformers",
     "url": "https://github.com/idiap/fast-transformers",
     "note": "PyTorch library for fast transformer implementations (1,785 stars)"
    },
    {
     "name": "importance-sampling",
     "url": "https://github.com/idiap/importance-sampling",
     "note": "Experiments on importance sampling for neural network training (330 stars)"
    },
    {
     "name": "bob",
     "url": "https://github.com/idiap/bob",
     "note": "Signal-processing and machine learning toolbox (262 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "Idiap/gated-deltanet-nsa-0.4B-10B",
     "url": "https://huggingface.co/Idiap/gated-deltanet-nsa-0.4B-10B",
     "kind": "model",
     "note": "Gated DeltaNet language model variant."
    },
    {
     "name": "Idiap/gated-deltanet-swa-1.4B-30B",
     "url": "https://huggingface.co/Idiap/gated-deltanet-swa-1.4B-30B",
     "kind": "model",
     "note": "Gated DeltaNet model with sliding-window attention."
    },
    {
     "name": "Idiap/EdgeFace",
     "url": "https://huggingface.co/spaces/Idiap/EdgeFace",
     "kind": "space",
     "note": "EdgeFace face recognition demo."
    },
    {
     "name": "Idiap/HyperFace",
     "url": "https://huggingface.co/spaces/Idiap/HyperFace",
     "kind": "space",
     "note": "HyperFace demo."
    }
   ],
   "papers": [
    {
     "title": "Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention",
     "url": "https://arxiv.org/abs/2006.16236",
     "year": "2020"
    },
    {
     "title": "Not All Samples Are Created Equal: Deep Learning with Importance Sampling",
     "url": "https://arxiv.org/abs/1803.00942",
     "year": "2018"
    },
    {
     "title": "ESLAM: Efficient Dense SLAM System Based on Hybrid Representation of Signed Distance Fields",
     "url": "https://arxiv.org/abs/2211.11704",
     "year": "2022"
    },
    {
     "title": "EdgeFace: Efficient Face Recognition Model for Edge Devices",
     "url": "https://arxiv.org/abs/2307.01838",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "fast-transformers",
     "url": "https://github.com/idiap/fast-transformers",
     "note": "PyTorch library for fast transformer implementations"
    },
    {
     "name": "Bob",
     "url": "https://github.com/idiap/bob",
     "note": "Signal-processing and machine learning toolbox"
    },
    {
     "name": "EdgeFace",
     "url": "https://huggingface.co/spaces/Idiap/EdgeFace",
     "note": "Face recognition demo"
    }
   ],
   "openness": "Large GitHub organisation; models and spaces are on Hugging Face.",
   "key_people": [
    {
     "name": "Sébastien Marcel",
     "role": "Member of the interim management team",
     "url": "https://www.idiap.ch/en/about",
     "note": "Named on the institute's own about page"
    },
    {
     "name": "Samy Bengio",
     "role": "Co-developer of Torch",
     "url": "https://en.wikipedia.org/wiki/Idiap_Research_Institute",
     "note": null
    },
    {
     "name": "Ronan Collobert",
     "role": "Co-developer of Torch",
     "url": "https://en.wikipedia.org/wiki/Idiap_Research_Institute",
     "note": null
    },
    {
     "name": "Johnny Mariéthoz",
     "role": "Co-developer of Torch",
     "url": "https://en.wikipedia.org/wiki/Idiap_Research_Institute",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The institute's contact page has a contact form and a secretariat line.",
   "sources": [
    "https://www.idiap.ch",
    "https://huggingface.co/Idiap",
    "https://github.com/idiap",
    "https://arxiv.org/abs/2006.16236"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "idiap-research-institute",
   "logo": true
  },
  {
   "orig_id": "illuin",
   "name": "Illuin Technology",
   "alias": null,
   "full_name": "Illuin Technology",
   "type": "startup",
   "founded": null,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.illuin.tech",
   "huggingface": "https://huggingface.co/illuin",
   "github": "https://github.com/illuin-tech",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Paris AI company whose research team co-created ColPali visual document retrieval and the ViDoRe benchmark, and co-authored the CroissantLLM bilingual model.",
   "tagline": "Illuin Technology is a Paris AI company whose research team co-created ColPali visual document retrieval and the ViDoRe benchmark.",
   "about": [
    "Illuin Technology is a Paris-based AI company that publishes research code under the illuin-tech GitHub organization and models on Hugging Face. Its researchers work with academic partners including CentraleSupélec and EPFL.",
    "In February 2024 the team co-authored CroissantLLM, a 1.3B-parameter bilingual French-English model with open training data. In June 2024 it released ColPali, which retrieves documents using vision-language models, together with the ViDoRe benchmark, and in 2025 it followed with the ModernVBERT encoder under the MIT license."
   ],
   "mission": "To build practical AI systems and advance open research on document understanding and retrieval.",
   "notice": null,
   "known_for": [
    "ColPali visual document retrieval",
    "ViDoRe benchmark",
    "CroissantLLM bilingual model"
   ],
   "focus": [
    "document retrieval",
    "vision-language models",
    "NLP",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "CroissantLLM",
     "note": "A 1.3B model pretrained on 3 trillion English and French tokens was released with open data and FrenchBench.",
     "year": "2024"
    },
    {
     "title": "ColPali",
     "note": "Faysse and colleagues introduced efficient document retrieval with vision language models.",
     "year": "2024"
    },
    {
     "title": "ViDoRe benchmark",
     "note": "Evaluation code for visual document retrieval was released alongside ColPali.",
     "year": "2024"
    },
    {
     "title": "Contextual embeddings",
     "note": "A model implementation for contextual embeddings was published.",
     "year": "2025"
    },
    {
     "title": "ModernVBERT",
     "note": "A 250M-parameter vision-language encoder for document retrieval was released under the MIT license.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "colpali",
     "url": "https://github.com/illuin-tech/colpali",
     "note": "The code used to train and run inference with the ColVision models, e.g. ColPali, ColQwen2, and ColSmol. (2,855 stars)"
    },
    {
     "name": "vidore-benchmark",
     "url": "https://github.com/illuin-tech/vidore-benchmark",
     "note": "Vision Document Retrieval (ViDoRe): Benchmark. Evaluation code for the ColPali paper. (278 stars)"
    },
    {
     "name": "opyoid",
     "url": "https://github.com/illuin-tech/opyoid",
     "note": "Dependency injection library for Python (75 stars)"
    },
    {
     "name": "contextual-embeddings",
     "url": "https://github.com/illuin-tech/contextual-embeddings",
     "note": "Model implementation for the contextual embeddings project (55 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "illuin/bert-large-uncased",
     "url": "https://huggingface.co/illuin/bert-large-uncased",
     "kind": "model",
     "note": "Model with 0 likes on the Hub."
    },
    {
     "name": "illuin/roberta-large-bne",
     "url": "https://huggingface.co/illuin/roberta-large-bne",
     "kind": "model",
     "note": "Model with 0 likes on the Hub."
    },
    {
     "name": "illuin/daedalus-traces",
     "url": "https://huggingface.co/datasets/illuin/daedalus-traces",
     "kind": "dataset",
     "note": "Dataset with 5 likes on the Hub."
    },
    {
     "name": "illuin/grouse",
     "url": "https://huggingface.co/spaces/illuin/grouse",
     "kind": "space",
     "note": "Space with 1 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "ColPali: Efficient Document Retrieval with Vision Language Models",
     "url": "https://arxiv.org/abs/2407.01449",
     "year": "2024"
    },
    {
     "title": "CroissantLLM: A Truly Bilingual French-English Language Model",
     "url": "https://arxiv.org/abs/2402.00786",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "ColPali / ColVision",
     "url": "https://github.com/illuin-tech/colpali",
     "note": "Code to train and run ColPali, ColQwen and related late-interaction vision retrievers"
    },
    {
     "name": "vidore-benchmark",
     "url": "https://github.com/illuin-tech/vidore-benchmark",
     "note": "Evaluation code for the ViDoRe visual document retrieval benchmark"
    },
    {
     "name": "GROUSE",
     "url": "https://huggingface.co/datasets/illuin/grouse",
     "note": "Dataset published on the Hub"
    },
    {
     "name": "opyoid",
     "url": "https://github.com/illuin-tech/opyoid",
     "note": "Python dependency injection library"
    }
   ],
   "openness": "Publishes research code, benchmarks and datasets under its illuin-tech GitHub organisation and illuin Hugging Face organisation; consulting and products are commercial.",
   "key_people": [
    {
     "name": "Manuel Faysse",
     "role": "ColPali and CroissantLLM first author",
     "url": "https://arxiv.org/abs/2407.01449",
     "note": null
    },
    {
     "name": "Gautier Viaud",
     "role": "ColPali and ModernVBERT co-author",
     "url": "https://arxiv.org/abs/2407.01449",
     "note": null
    },
    {
     "name": "Hugues Sibille",
     "role": "ColPali co-author",
     "url": "https://arxiv.org/abs/2407.01449",
     "note": null
    },
    {
     "name": "Tony Wu",
     "role": "ColPali co-author",
     "url": "https://arxiv.org/abs/2407.01449",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://www.illuin.tech",
    "https://huggingface.co/illuin",
    "https://github.com/illuin-tech",
    "https://arxiv.org/abs/2407.01449"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "illuin-technology",
   "mlat": 48.91872,
   "mlon": 2.47722,
   "logo": true
  },
  {
   "orig_id": "ilsp-athena",
   "name": "ILSP / Athena Research Center",
   "alias": "Meltemi, Krikri",
   "full_name": "ILSP / Athena Research Center (Meltemi, Krikri)",
   "type": "institute",
   "founded": null,
   "country": "Greece",
   "city": "Athens",
   "lat": 37.9838,
   "lon": 23.7275,
   "website": "https://www.athenarc.gr",
   "huggingface": "https://huggingface.co/ilsp",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Institute for Language and Speech Processing of Athena Research Center, which builds open Greek language models such as Meltemi and Llama-Krikri.",
   "tagline": "Greek language technology institute at Athena Research Center that released Meltemi, the first open large language model for Greek, and Llama-Krikri.",
   "about": [
    "The Institute for Language and Speech Processing (ILSP) is part of Athena Research Center in Athens and works on language and speech technology. It has recently marked 35 years of activity. Its research covers language resources, speech processing and, more recently, large language models for Greek.",
    "In March 2024 the institute released Meltemi 7B, trained on a corpus of about 40 billion Greek tokens, and described as the first open large language model for Greek in a July 2024 paper. It followed with Llama-Krikri-8B, built on Llama 3.1 with an extended Greek tokenizer and a 128k context window, trained on about 56.7 billion Greek tokens and supporting polytonic and Ancient Greek.",
    "Models and instruction-tuned variants are published on Hugging Face under the ilsp organization, giving Greek developers and researchers open alternatives to general multilingual models."
   ],
   "mission": "ILSP develops language and speech technology for Greek and openly releases the resulting models and resources so that the language is well supported in modern AI systems.",
   "notice": null,
   "known_for": [
    "Meltemi, the first open Greek LLM",
    "Llama-Krikri Greek models",
    "Greek language and speech resources"
   ],
   "focus": [
    "Greek",
    "language models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Meltemi 7B",
     "note": "Released the first open large language model for Greek, trained on about 40 billion Greek tokens.",
     "year": "2024"
    },
    {
     "title": "Meltemi paper",
     "note": "Published a technical paper describing Meltemi and its evaluation.",
     "year": "2024"
    },
    {
     "title": "Llama-Krikri-8B",
     "note": "Released a Greek foundation model based on Llama 3.1 with Greek vocabulary extension and 128k context.",
     "year": "2025"
    },
    {
     "title": "Krikri Instruct v1.5",
     "note": "Published an instruction-tuned update of the Krikri family.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "ilsp/Llama-Krikri-8B-Instruct",
     "url": "https://huggingface.co/ilsp/Llama-Krikri-8B-Instruct",
     "kind": "model",
     "note": "Greek instruct model, the most downloaded ILSP model."
    },
    {
     "name": "ilsp/Llama-Krikri-8B-Base",
     "url": "https://huggingface.co/ilsp/Llama-Krikri-8B-Base",
     "kind": "model",
     "note": "Greek base model."
    },
    {
     "name": "ilsp/mmlu_greek",
     "url": "https://huggingface.co/datasets/ilsp/mmlu_greek",
     "kind": "dataset",
     "note": "Greek MMLU evaluation set."
    },
    {
     "name": "ilsp/Meltemi-7B-v1",
     "url": "https://huggingface.co/ilsp/Meltemi-7B-v1",
     "kind": "model",
     "note": "First open Greek LLM."
    }
   ],
   "papers": [
    {
     "title": "Meltemi: The first open Large Language Model for Greek",
     "url": "https://arxiv.org/abs/2407.20743",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Meltemi-7B-v1",
     "url": "https://huggingface.co/ilsp/Meltemi-7B-v1",
     "note": "First open Greek LLM"
    },
    {
     "name": "Llama-Krikri-8B-Instruct",
     "url": "https://huggingface.co/ilsp/Llama-Krikri-8B-Instruct",
     "note": "Greek instruction model"
    },
    {
     "name": "hellaswag_greek",
     "url": "https://huggingface.co/datasets/ilsp/hellaswag_greek",
     "note": "Greek evaluation dataset"
    }
   ],
   "openness": "Models and evaluation datasets are public on Hugging Face.",
   "key_people": [
    {
     "name": "Vassilis Katsouros",
     "role": "Co-author of the Meltemi paper",
     "url": "https://arxiv.org/abs/2407.20743",
     "note": null
    },
    {
     "name": "Leon Voukoutis",
     "role": "Co-author of the Meltemi paper",
     "url": "https://arxiv.org/abs/2407.20743",
     "note": null
    },
    {
     "name": "Dimitris Roussis",
     "role": "Co-author of the Meltemi paper",
     "url": "https://arxiv.org/abs/2407.20743",
     "note": null
    },
    {
     "name": "Georgios Paraskevopoulos",
     "role": "Co-author of the Meltemi paper",
     "url": "https://arxiv.org/abs/2407.20743",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details sit on the Athena Research Center website; the page could not be fetched during checking.",
   "sources": [
    "https://www.athenarc.gr",
    "https://huggingface.co/ilsp",
    "https://arxiv.org/abs/2407.20743"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "ilsp-athena-research-center",
   "logo": true
  },
  {
   "orig_id": "insait",
   "name": "INSAIT",
   "alias": null,
   "full_name": "INSAIT",
   "type": "institute",
   "founded": 2022,
   "country": "Bulgaria",
   "city": "Sofia",
   "lat": 42.6977,
   "lon": 23.3219,
   "website": "https://insait.ai",
   "huggingface": "https://huggingface.co/INSAIT-Institute",
   "github": "https://github.com/insait-institute",
   "x": "https://x.com/INSAITinstitute",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/insaitinstitute",
   "contact_email": null,
   "summary": "AI and computer science research institute at Sofia University. It publishes the BgGPT and MamayLM language models and computer-vision research code.",
   "tagline": "Bulgarian AI and computer science institute at Sofia University, in partnership with ETH Zurich and EPFL, behind the BgGPT and MamayLM open language models.",
   "about": [
    "INSAIT, the Institute for Computer Science, Artificial Intelligence and Technology, was created by decree of the Bulgarian Council of Ministers in February 2021 and began operations in April 2022. It is hosted at Sofia University and set up in partnership with ETH Zurich, EPFL and the Government of Bulgaria. Its founder and scientific director is Martin Vechev, a professor at ETH Zurich.",
    "The Bulgarian government committed about 85 million euros over ten years, and donors include Google, AWS, DeepMind, SiteGround and VMware. Besides foundation models, the institute runs research in computer vision, robotics, security and formal methods, and spins out companies such as LogicStar AI.",
    "Its open model line includes BgGPT for Bulgarian and MamayLM for Ukrainian, which was developed with ETH Zurich on Gemma 2 9B. In 2024 it also released COMPL-AI, a framework for evaluating models against the EU AI Act."
   ],
   "mission": "INSAIT states its mission as achieving excellence in science, research and education, and it applies this to building language technology for under-served European languages.",
   "notice": null,
   "known_for": [
    "BgGPT, open Bulgarian language models",
    "MamayLM for Ukrainian",
    "COMPL-AI EU AI Act evaluation framework"
   ],
   "focus": [
    "language models",
    "Bulgarian",
    "Ukrainian",
    "vision"
   ],
   "contributions": [
    {
     "title": "Institute created by decree",
     "note": "The Bulgarian Council of Ministers formally created INSAIT with ETH Zurich, EPFL and Sofia University.",
     "year": "2021"
    },
    {
     "title": "Operations begin",
     "note": "INSAIT officially began operations in Sofia.",
     "year": "2022"
    },
    {
     "title": "AI Alliance founding member",
     "note": "Joined the AI Alliance as a founding member.",
     "year": "2023"
    },
    {
     "title": "COMPL-AI",
     "note": "Released an evaluation framework for EU AI Act compliance with ETH Zurich and LatticeFlow AI.",
     "year": "2024"
    },
    {
     "title": "BgGPT public release",
     "note": "Published the BgGPT open models for Bulgarian.",
     "year": "2024"
    },
    {
     "title": "MamayLM",
     "note": "Released the Ukrainian model MamayLM, co-developed with ETH Zurich.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "GenieRedux",
     "url": "https://github.com/insait-institute/GenieRedux",
     "note": "A framework for training world models with virtual environments, complete with annotated environment dataset ( (79 stars)"
    },
    {
     "name": "OccamLGS",
     "url": "https://github.com/insait-institute/OccamLGS",
     "note": "[BMVC 2025] Occam’s LGS: An Efficient Approach for Language Gaussian Splatting (73 stars)"
    },
    {
     "name": "USDNet",
     "url": "https://github.com/insait-institute/USDNet",
     "note": "Open-source repository (34 stars)"
    },
    {
     "name": "AR-VLA-lerobot",
     "url": "https://github.com/insait-institute/AR-VLA-lerobot",
     "note": "AR actor for specialist policy training (34 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "INSAIT-Institute/BgGPT-Gemma-3-27B-IT-GGUF",
     "url": "https://huggingface.co/INSAIT-Institute/BgGPT-Gemma-3-27B-IT-GGUF",
     "kind": "model",
     "note": "Model: 2,878 downloads, 8 likes."
    },
    {
     "name": "INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v2.0-GGUF",
     "url": "https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v2.0-GGUF",
     "kind": "model",
     "note": "Model: 1,895 downloads, 5 likes."
    },
    {
     "name": "INSAIT-Institute/Articulate3D",
     "url": "https://huggingface.co/datasets/INSAIT-Institute/Articulate3D",
     "kind": "dataset",
     "note": "Dataset: 3,556 downloads, 4 likes."
    },
    {
     "name": "INSAIT-Institute/mamaylm-v1-blog",
     "url": "https://huggingface.co/spaces/INSAIT-Institute/mamaylm-v1-blog",
     "kind": "space",
     "note": "Space: 27 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "BgGPT 1.0: Extending English-centric LLMs to other languages",
     "url": "https://arxiv.org/abs/2412.10893",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "BgGPT-7B-Instruct v0.2",
     "url": "https://huggingface.co/INSAIT-Institute/BgGPT-7B-Instruct-v0.2",
     "note": "Bulgarian-adapted language model"
    },
    {
     "name": "MamayLM",
     "url": "https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-2-9B-IT-v0.1",
     "note": "Ukrainian-adapted language model"
    },
    {
     "name": "GenieRedux",
     "url": "https://github.com/insait-institute/GenieRedux",
     "note": "World-model training framework"
    }
   ],
   "openness": "Models and code are public on Hugging Face and GitHub; models adapted from third-party bases carry those licences.",
   "key_people": [
    {
     "name": "Martin Vechev",
     "role": "Founder and Scientific Director",
     "url": "https://insait.ai",
     "note": null
    },
    {
     "name": "Borislav Petrov",
     "role": "Executive Director",
     "url": "https://en.wikipedia.org/wiki/INSAIT",
     "note": null
    },
    {
     "name": "Virginia Vassilevska Williams",
     "role": "Professor listed on INSAIT's site",
     "url": "https://insait.ai",
     "note": null
    },
    {
     "name": "Andreas Krause",
     "role": "Professor listed on INSAIT's site",
     "url": "https://insait.ai",
     "note": null
    },
    {
     "name": "Srdjan Capkun",
     "role": "Professor listed on INSAIT's site",
     "url": "https://insait.ai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the institute website.",
   "sources": [
    "https://insait.ai",
    "https://huggingface.co/INSAIT-Institute",
    "https://github.com/insait-institute"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "insait",
   "logo": true
  },
  {
   "orig_id": "ml-jku",
   "name": "Institute for Machine Learning, JKU Linz",
   "alias": null,
   "full_name": "Institute for Machine Learning, JKU Linz",
   "type": "academic",
   "founded": null,
   "country": "Austria",
   "city": "Linz",
   "lat": 48.3069,
   "lon": 14.2858,
   "website": "https://www.jku.at/en/institute-for-machine-learning/",
   "huggingface": "https://huggingface.co/ml-jku",
   "github": "https://github.com/ml-jku",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Machine learning institute at Johannes Kepler University Linz, known for LSTM, modern Hopfield networks and xLSTM.",
   "tagline": "Johannes Kepler University's machine learning institute, led by Sepp Hochreiter, the inventor of LSTM and a driver of modern Hopfield networks and xLSTM.",
   "about": [
    "The Institute for Machine Learning at Johannes Kepler University Linz has been led since 2018 by Sepp Hochreiter. Hochreiter developed the long short-term memory architecture in his 1991 diploma thesis and published it with Jürgen Schmidhuber in 1997. He previously led the university's Institute of Bioinformatics and heads the LIT AI Lab, which he took over in 2017.",
    "The group works on machine learning and statistical methods. Its researchers introduced modern Hopfield networks with continuous states in 2020 and applied them to immune repertoire classification. In 2024 the group and collaborators published xLSTM, an extended LSTM architecture presented at NeurIPS, and the company NXAI, where Hochreiter is chief scientist, builds industrial models on it.",
    "Code and models appear on the institute's GitHub and Hugging Face organizations."
   ],
   "mission": "The institute develops and applies machine learning and statistical methods, with a stated emphasis on deep learning fundamentals and their use in science and industry.",
   "notice": null,
   "known_for": [
    "Invention of LSTM",
    "Modern Hopfield networks",
    "xLSTM architecture"
   ],
   "focus": [
    "architectures",
    "science",
    "reinforcement learning"
   ],
   "contributions": [
    {
     "title": "LSTM",
     "note": "Hochreiter and Schmidhuber published the long short-term memory architecture in Neural Computation.",
     "year": "1997"
    },
    {
     "title": "LIT AI Lab",
     "note": "Hochreiter became head of the Linz Institute of Technology AI Lab.",
     "year": "2017"
    },
    {
     "title": "Institute for Machine Learning",
     "note": "Hochreiter began leading the Institute for Machine Learning at JKU.",
     "year": "2018"
    },
    {
     "title": "Modern Hopfield networks",
     "note": "Ramsauer and colleagues introduced Hopfield networks with continuous states.",
     "year": "2020"
    },
    {
     "title": "xLSTM",
     "note": "Published Extended LSTM at NeurIPS as a scalable alternative to Transformers and state space models.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "hopfield-layers",
     "url": "https://github.com/ml-jku/hopfield-layers",
     "note": "Hopfield Networks is All You Need (1,977 stars)"
    },
    {
     "name": "hopular",
     "url": "https://github.com/ml-jku/hopular",
     "note": "Modern Hopfield networks for tabular data (317 stars)"
    },
    {
     "name": "UPT",
     "url": "https://github.com/ml-jku/UPT",
     "note": "Code for Universal Physics Transformers (179 stars)"
    },
    {
     "name": "GINNs-Geometry-informed-Neural-Networks",
     "url": "https://github.com/ml-jku/GINNs-Geometry-informed-Neural-Networks",
     "note": "Geometry-informed neural networks (195 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "ml-jku/dm_control",
     "url": "https://huggingface.co/datasets/ml-jku/dm_control",
     "kind": "dataset",
     "note": "Control task data for offline reinforcement learning."
    },
    {
     "name": "ml-jku/dark_room",
     "url": "https://huggingface.co/datasets/ml-jku/dark_room",
     "kind": "dataset",
     "note": "Dark-room reinforcement learning data."
    },
    {
     "name": "ml-jku/mhnfs",
     "url": "https://huggingface.co/spaces/ml-jku/mhnfs",
     "kind": "space",
     "note": "Few-shot molecule property prediction with modern Hopfield networks."
    },
    {
     "name": "ml-jku/tox21_leaderboard",
     "url": "https://huggingface.co/spaces/ml-jku/tox21_leaderboard",
     "kind": "space",
     "note": "Tox21 toxicity prediction leaderboard."
    }
   ],
   "papers": [
    {
     "title": "Hopfield Networks is All You Need",
     "url": "https://arxiv.org/abs/2008.02217",
     "year": "2020"
    },
    {
     "title": "RUDDER: Return Decomposition for Delayed Rewards",
     "url": "https://arxiv.org/abs/1806.07857",
     "year": "2018"
    },
    {
     "title": "Self-Normalizing Neural Networks",
     "url": "https://arxiv.org/abs/1706.02515",
     "year": "2017"
    },
    {
     "title": "Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators",
     "url": "https://arxiv.org/abs/2402.12365",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "hopfield-layers",
     "url": "https://github.com/ml-jku/hopfield-layers",
     "note": "Hopfield network layers for deep learning"
    },
    {
     "name": "UPT",
     "url": "https://github.com/ml-jku/UPT",
     "note": "Universal Physics Transformers"
    },
    {
     "name": "hopular",
     "url": "https://github.com/ml-jku/hopular",
     "note": "Hopfield networks for tabular data"
    }
   ],
   "openness": "Code is public on GitHub and datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Sepp Hochreiter",
     "role": "Head of the institute",
     "url": "https://en.wikipedia.org/wiki/Sepp_Hochreiter",
     "note": null
    },
    {
     "name": "Günter Klambauer",
     "role": "Researcher, Hopfield and xLSTM co-author",
     "url": "https://www.jku.at/en/institute-for-machine-learning/",
     "note": null
    },
    {
     "name": "Maximilian Beck",
     "role": "First author of xLSTM",
     "url": "https://proceedings.neurips.cc/paper_files/paper/2024/file/c2ce2f2701c10a2b2f2ea0bfa43cfaa3-Paper-Conference.pdf",
     "note": null
    },
    {
     "name": "Korbinian Pöppel",
     "role": "Co-author of xLSTM",
     "url": "https://proceedings.neurips.cc/paper_files/paper/2024/file/c2ce2f2701c10a2b2f2ea0bfa43cfaa3-Paper-Conference.pdf",
     "note": null
    },
    {
     "name": "Johannes Brandstetter",
     "role": "Co-author of xLSTM",
     "url": "https://proceedings.neurips.cc/paper_files/paper/2024/file/c2ce2f2701c10a2b2f2ea0bfa43cfaa3-Paper-Conference.pdf",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details are given on the institute's pages of the JKU Linz website.",
   "sources": [
    "https://www.jku.at/en/institute-for-machine-learning/",
    "https://huggingface.co/ml-jku",
    "https://github.com/ml-jku",
    "https://arxiv.org/abs/2008.02217"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "institute-for-machine-learning-jku-linz",
   "mlat": 48.27154,
   "mlon": 14.2858,
   "logo": true
  },
  {
   "orig_id": "ipipan",
   "name": "Institute of Computer Science PAS",
   "alias": "IPI PAN",
   "full_name": "Institute of Computer Science PAS (IPI PAN)",
   "type": "institute",
   "founded": null,
   "country": "Poland",
   "city": "Warsaw",
   "lat": 52.2297,
   "lon": 21.0122,
   "website": "https://ipipan.waw.pl",
   "huggingface": "https://huggingface.co/ipipan",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Institute of Computer Science of the Polish Academy of Sciences, publishing open Polish retrieval models and question answering datasets.",
   "tagline": "IPI PAN, the Institute of Computer Science of the Polish Academy of Sciences, publishes open Polish retrieval models and question answering datasets.",
   "about": [
    "The Institute of Computer Science of the Polish Academy of Sciences in Warsaw hosts a Linguistic Engineering Group that builds resources for Polish. Its researchers co-trained HerBERT with Allegro, and Alina Wróblewska is a co-author of that work.",
    "Its Hugging Face organization lists dozens of models and datasets, including the PolQA and MAUPQA question answering datasets, the NKJP1M corpus, the Silver Retriever retrieval model, and HerBERT variants for Silesian and Kashubian."
   ],
   "mission": "To provide open language resources and models for Polish and related regional languages.",
   "notice": null,
   "known_for": [
    "Polish question answering datasets PolQA and MAUPQA",
    "Silver Retriever Polish retrieval model",
    "HerBERT co-development"
   ],
   "focus": [
    "Polish NLP",
    "retrieval"
   ],
   "contributions": [
    {
     "title": "NKJP1M and herference",
     "note": "The NKJP1M corpus and the herference-large model were published on Hugging Face.",
     "year": "2022"
    },
    {
     "title": "PolQA",
     "note": "A Polish question answering dataset was released under CC BY-SA 4.0.",
     "year": "2022"
    },
    {
     "title": "MAUPQA",
     "note": "A large Polish question answering dataset was released.",
     "year": "2023"
    },
    {
     "title": "Silver Retriever",
     "note": "A Polish passage retrieval model was released.",
     "year": "2023"
    },
    {
     "title": "Silesian and Kashubian HerBERT",
     "note": "Regional language models based on HerBERT were published.",
     "year": "2023"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "ipipan/silver-retriever-base-v1",
     "url": "https://huggingface.co/ipipan/silver-retriever-base-v1",
     "kind": "model",
     "note": "Model with 14 likes on the Hub."
    },
    {
     "name": "ipipan/silver-retriever-base-v1.1",
     "url": "https://huggingface.co/ipipan/silver-retriever-base-v1.1",
     "kind": "model",
     "note": "Model with 12 likes on the Hub."
    },
    {
     "name": "ipipan/polqa",
     "url": "https://huggingface.co/datasets/ipipan/polqa",
     "kind": "dataset",
     "note": "Dataset with 13 likes on the Hub."
    },
    {
     "name": "ipipan/nlprepl",
     "url": "https://huggingface.co/spaces/ipipan/nlprepl",
     "kind": "space",
     "note": "Space with 4 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Silver Retriever: Advancing Neural Passage Retrieval for Polish Question Answering",
     "url": "https://arxiv.org/abs/2309.08469",
     "year": "2023"
    },
    {
     "title": "PolQA: Polish Question Answering Dataset",
     "url": "https://arxiv.org/abs/2212.08897",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "silver-retriever-base-v1",
     "url": "https://huggingface.co/ipipan/silver-retriever-base-v1",
     "note": "Polish passage retriever"
    },
    {
     "name": "PolQA",
     "url": "https://huggingface.co/datasets/ipipan/polqa",
     "note": "Polish question answering dataset"
    },
    {
     "name": "MAUPQA",
     "url": "https://huggingface.co/datasets/ipipan/maupqa",
     "note": "Polish passage retrieval dataset"
    },
    {
     "name": "NKJP1M",
     "url": "https://huggingface.co/datasets/ipipan/nkjp1m",
     "note": "Annotated Polish corpus"
    }
   ],
   "openness": "Models and datasets are published on the ipipan Hugging Face organisation.",
   "key_people": [
    {
     "name": "Alina Wróblewska",
     "role": "Researcher, HerBERT co-author",
     "url": "https://huggingface.co/alina-wroblewska",
     "note": null
    },
    {
     "name": "Agnieszka Mykowiecka",
     "role": "Researcher",
     "url": "https://huggingface.co/amykowiecka",
     "note": null
    },
    {
     "name": "Aleksander Wawer",
     "role": "Researcher",
     "url": "https://huggingface.co/aleksanderwawer",
     "note": null
    },
    {
     "name": "Piotr Rychlik",
     "role": "Researcher",
     "url": "https://huggingface.co/rychlik",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://ipipan.waw.pl",
    "https://huggingface.co/ipipan",
    "https://arxiv.org/abs/2309.08469"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "institute-of-computer-science-pas",
   "mlat": 52.19434,
   "mlon": 21.0122,
   "logo": true
  },
  {
   "orig_id": "institute-of-foundation-models",
   "name": "Institute of Foundation Models",
   "alias": "LLM360",
   "full_name": "Institute of Foundation Models (LLM360)",
   "type": "institute",
   "founded": null,
   "country": "United Arab Emirates",
   "city": "Abu Dhabi",
   "lat": 24.45,
   "lon": 54.38,
   "website": "https://www.llm360.ai",
   "huggingface": "https://huggingface.co/LLM360",
   "github": "https://github.com/LLM360",
   "x": null,
   "discord": "https://discord.gg/jFPq9AycBu",
   "linkedin": null,
   "contact_email": null,
   "summary": "Research institute at MBZUAI in Abu Dhabi that sponsors LLM360, the open model and data effort. The legacy LLM360 Hugging Face organization notes that its namespace has moved to the Institute of Foundation Models.",
   "tagline": "MBZUAI research institute that sponsors LLM360, the fully open effort behind the Amber, K2 and TxT360 releases, and the K2 Horizon model family.",
   "about": [
    "The Institute of Foundation Models (IFM) is a research unit of the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi, launched publicly in May 2025. The LLM360 project, which states it is proudly sponsored by the institute, publishes open large language models together with their training data, code and checkpoints. The LLM360 Hugging Face organization notes that its namespace has moved to IFM.",
    "LLM360's releases include Amber-7B, Crystal-7B, the 65-billion-parameter K2 trained on 1.4 trillion tokens, and TxT360, a pre-training dataset that deduplicates 99 Common Crawl snapshots and 14 other sources. In September 2026 the institute released K2 Horizon, six fully open models from 0.9 to 375 billion parameters.",
    "The institute also coordinates the Jais Arabic models and operates with a Silicon Valley lab alongside its Abu Dhabi base."
   ],
   "mission": "LLM360 aims to enable community-owned AI through open large model research, creating standards and tools for capability and knowledge transfer. The institute releases weights, code, data and training details.",
   "notice": null,
   "known_for": [
    "LLM360 fully open model releases",
    "K2 and K2 Horizon models",
    "TxT360 pre-training dataset"
   ],
   "focus": [
    "language models",
    "training transparency"
   ],
   "contributions": [
    {
     "title": "Amber-7B and Crystal-7B",
     "note": "Released open 7B models with full training artifacts and chat variants.",
     "year": "2023"
    },
    {
     "title": "K2-65B",
     "note": "Released a 65B model trained on 1.4 trillion tokens that LLM360 reports matches Llama 2 70B with less compute.",
     "year": "2024"
    },
    {
     "title": "TxT360",
     "note": "Released a pre-training dataset globally deduplicating 99 Common Crawl snapshots and 14 other sources.",
     "year": "2024"
    },
    {
     "title": "Institute launch",
     "note": "MBZUAI launched the Institute of Foundation Models and a Silicon Valley lab.",
     "year": "2025"
    },
    {
     "title": "K2 Horizon",
     "note": "Released six fully open models from 0.9B to 375B parameters with data and methodology.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "amber-train",
     "url": "https://github.com/LLM360/amber-train",
     "note": "Pretraining code for Amber 7B."
    },
    {
     "name": "Reasoning360",
     "url": "https://github.com/LLM360/Reasoning360",
     "note": "Open research on large reasoning models."
    },
    {
     "name": "Analysis360",
     "url": "https://github.com/LLM360/Analysis360",
     "note": "Open LLM analysis implementations."
    },
    {
     "name": "MegaMath",
     "url": "https://github.com/LLM360/MegaMath",
     "note": "Open 370B-token math pretraining dataset."
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "LLM360: Towards Fully Transparent Open-Source LLMs",
     "url": "https://arxiv.org/abs/2312.06550",
     "year": "2023"
    },
    {
     "title": "K2-Think: A Parameter-Efficient Reasoning System",
     "url": "https://arxiv.org/abs/2509.07604",
     "year": "2025"
    }
   ],
   "notable_releases": [],
   "openness": "LLM360 is known for releasing training artefacts; the legacy HF org shows no public models.",
   "key_people": [
    {
     "name": "Eric Xing",
     "role": "President of MBZUAI; Professor, Carnegie Mellon University",
     "url": "https://www.cs.cmu.edu/~epxing/",
     "note": "Leads the university that hosts the Institute of Foundation Models and commented on K2 Think."
    }
   ],
   "discord_note": "Discord linked from llm360.ai for the LLM360 community.",
   "contact_note": "llm360.ai lists Twitter, Discord, GitHub and email links.",
   "sources": [
    "https://www.llm360.ai"
   ],
   "confidence": "medium",
   "region": "Middle East",
   "id": "institute-of-foundation-models",
   "mlat": 24.4067,
   "mlon": 54.38,
   "logo": true
  },
  {
   "orig_id": "issai",
   "name": "ISSAI",
   "alias": "Nazarbayev University",
   "full_name": "ISSAI (Nazarbayev University)",
   "type": "institute",
   "founded": 2019,
   "country": "Kazakhstan",
   "city": "Astana",
   "lat": 51.17,
   "lon": 71.45,
   "website": "http://issai.nu.edu.kz",
   "huggingface": "https://huggingface.co/issai",
   "github": "https://github.com/IS2AI",
   "x": null,
   "discord": null,
   "linkedin": "https://linkedin.com/company/institute-of-smart-systems-and-artificial-intelligence",
   "contact_email": "issai@nu.edu.kz",
   "summary": "AI institute at Nazarbayev University working on Kazakh-language AI.",
   "tagline": "Nazarbayev University's institute for AI in Kazakhstan, publishing open Kazakh speech corpora, TTS data and the KazLLM language model.",
   "about": [
    "The Institute of Smart Systems and Artificial Intelligence (ISSAI) was founded in September 2019 at Nazarbayev University in Astana. It was set up to build national research capacity in artificial intelligence and draws on experience from institutions in Asia, Europe and North America. Its home is the university's campus on Kabanbay Batyr Avenue.",
    "ISSAI is known for open resources for the Kazakh language, including the Kazakh Speech Corpus, whose second edition holds about 1,200 hours of transcribed speech with Kazakh-Russian code-switching, and the KazakhTTS corpora. It worked with the Ministry of Science and Higher Education on KazLLM, whose open-source version is on Hugging Face.",
    "The institute also runs applied projects such as the Tilsync partnership offering and publishes code on GitHub under the IS2AI organization."
   ],
   "mission": "ISSAI aims to develop a national capacity for artificial intelligence research in Kazakhstan, with a focus on the Kazakh language and open resources.",
   "notice": null,
   "known_for": [
    "Kazakh Speech Corpus",
    "KazakhTTS corpora",
    "KazLLM"
   ],
   "focus": [
    "Kazakh",
    "speech",
    "language models"
   ],
   "contributions": [
    {
     "title": "ISSAI founded",
     "note": "Established at Nazarbayev University to build national AI research capacity.",
     "year": "2019"
    },
    {
     "title": "Kazakh Speech Corpus",
     "note": "Published a crowdsourced open Kazakh speech corpus with baseline recognition results.",
     "year": "2021"
    },
    {
     "title": "KazakhTTS2",
     "note": "Expanded the open Kazakh TTS corpus from 93 to 271 hours with five speakers.",
     "year": "2022"
    },
    {
     "title": "KSC2",
     "note": "Released the second Kazakh Speech Corpus of roughly 1,200 hours with code-switched speech.",
     "year": "2022"
    },
    {
     "title": "KazLLM",
     "note": "Contributed to a Kazakh large language model built with the Ministry, with an open version on Hugging Face.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "Kazakh_TTS",
     "url": "https://github.com/IS2AI/Kazakh_TTS",
     "note": "Kazakh TTS."
    },
    {
     "name": "SpeakingFaces",
     "url": "https://github.com/IS2AI/SpeakingFaces",
     "note": "Visual-thermal-audio dataset."
    },
    {
     "name": "TurkicASR",
     "url": "https://github.com/IS2AI/TurkicASR",
     "note": "ASR for ten Turkic languages."
    },
    {
     "name": "TurkicTTS",
     "url": "https://github.com/IS2AI/TurkicTTS",
     "note": "TTS for ten Turkic languages."
    }
   ],
   "hf_highlights": [
    {
     "name": "issai/Qolda_GGUF",
     "url": "https://huggingface.co/issai/Qolda_GGUF",
     "kind": "model",
     "note": "Qolda model."
    },
    {
     "name": "issai/Qolda-AVL-5B-GGUF",
     "url": "https://huggingface.co/issai/Qolda-AVL-5B-GGUF",
     "kind": "model",
     "note": "Vision-language."
    },
    {
     "name": "issai/Apples_HSI",
     "url": "https://huggingface.co/datasets/issai/Apples_HSI",
     "kind": "dataset",
     "note": "Hyperspectral apples."
    },
    {
     "name": "issai/Kazakh_Speech_Corpus_2",
     "url": "https://huggingface.co/datasets/issai/Kazakh_Speech_Corpus_2",
     "kind": "dataset",
     "note": "Kazakh speech."
    }
   ],
   "papers": [
    {
     "title": "KazakhTTS: An Open-Source Kazakh Text-to-Speech Synthesis Dataset",
     "url": "https://arxiv.org/abs/2104.08459",
     "year": "2021"
    },
    {
     "title": "A Crowdsourced Open-Source Kazakh Speech Corpus and Initial Speech Recognition Baseline",
     "url": "https://arxiv.org/abs/2009.10334",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "foggen",
     "url": "https://huggingface.co/issai/foggen",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Models on Hugging Face and code on GitHub.",
   "key_people": [
    {
     "name": "Huseyin Atakan Varol",
     "role": "Founding and General Director",
     "url": "https://issai.nu.edu.kz/2021/01/20/director-3/",
     "note": "Leads the voice-technology team and co-authored the Kazakh speech and TTS corpora."
    },
    {
     "name": "Yerbol Absalyamov",
     "role": "Executive Director",
     "url": "https://issai.nu.edu.kz/2021/01/22/yerbol-absalyamov/",
     "note": "Listed on the institute's team page."
    },
    {
     "name": "Madina Abdrakhmanova",
     "role": "Deputy Director, Product and External Affairs; Senior Data Scientist",
     "url": "https://issai.nu.edu.kz/2021/01/25/madina-abdrakhmanova/",
     "note": "Listed on the institute's team page."
    },
    {
     "name": "Askat Kuzdeuov",
     "role": "Lead Data Scientist",
     "url": "https://issai.nu.edu.kz/2021/01/28/senior-researcher/",
     "note": "Listed on the institute's team page."
    }
   ],
   "discord_note": null,
   "contact_note": "issai.nu.edu.kz lists a general institute email.",
   "sources": [
    "https://huggingface.co/issai",
    "http://issai.nu.edu.kz"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "issai",
   "logo": true
  },
  {
   "orig_id": "jerteh",
   "name": "JeRTeh",
   "alias": "Society for Language Resources and Technologies",
   "full_name": "JeRTeh (Society for Language Resources and Technologies)",
   "type": "nonprofit",
   "founded": null,
   "country": "Serbia",
   "city": "Belgrade",
   "lat": 44.7866,
   "lon": 20.4489,
   "website": "https://jerteh.rs",
   "huggingface": "https://huggingface.co/jerteh",
   "github": "https://github.com/jerteh",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Serbian society for language resources and technologies that releases Serbian GPT-2 and BERT-style models and corpora.",
   "tagline": "JeRTeh is a Serbian society that releases Serbian language models, corpora and tools for language technology.",
   "about": [
    "The Society for Language Resources and Technologies, known as JeRTeh, is a Serbian association that runs seminars, publishes resources and takes part in projects such as UniDive, Distant Reading, WikiELTeC and It-Sr-NER. It maintains a GitHub organization opened in 2023.",
    "It publishes Serbian models such as Jerteh-355 on Hugging Face and corpora including srpELTeC. In April 2026 it joined a protocol with the Office for IT and eGovernment and ten other institutions to build a national Serbian language model."
   ],
   "mission": "To develop and share language resources and models for Serbian.",
   "notice": null,
   "known_for": [
    "Serbian GPT-2 and BERT-style models",
    "srpELTeC literary corpus",
    "Serbian language technology society"
   ],
   "focus": [
    "Serbian NLP",
    "language models",
    "corpora"
   ],
   "contributions": [
    {
     "title": "BEaSTagger",
     "note": "A stacked POS tagging system for Serbian was published.",
     "year": "2021"
    },
    {
     "title": "It-Sr-NER",
     "note": "CLARIN-compatible NER and geoparsing services for Italian and Serbian were released.",
     "year": "2022"
    },
    {
     "title": "Jerteh-355 model",
     "note": "A Serbian language model was published on Hugging Face.",
     "year": "2023"
    },
    {
     "title": "NovaIskra project",
     "note": "A cultural heritage project with UNILIB was started.",
     "year": "2024"
    },
    {
     "title": "National language model protocol",
     "note": "JeRTeh and eleven institutions signed a protocol to build a Serbian national model.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "BEaSTagger",
     "url": "https://github.com/jerteh/BEaSTagger",
     "note": "Bidirectional, Expandable and Stacked Tagger --- BEaST. This package provides POS-tagging using a system compo (2 stars)"
    },
    {
     "name": "It-Sr-NER",
     "url": "https://github.com/jerteh/It-Sr-NER",
     "note": "It-Sr-NER: CLARIN compatible NER and geoparsing web services  for Italian and Serbian parallel text (2 stars)"
    },
    {
     "name": "NovaIskra",
     "url": "https://github.com/jerteh/NovaIskra",
     "note": "UNILIB & JeRTeh :: projekat u oblasti kulturnog nasleđa i bibliotečko-informacione delatnosti  (0 stars)"
    },
    {
     "name": "srpELTeC-ijk",
     "url": "https://github.com/jerteh/srpELTeC-ijk",
     "note": "Serbian ELTeC ijekavijen collection (0 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "jerteh/gpt2-orao",
     "url": "https://huggingface.co/jerteh/gpt2-orao",
     "kind": "model",
     "note": "Model with 13 likes on the Hub."
    },
    {
     "name": "jerteh/Jerteh-355",
     "url": "https://huggingface.co/jerteh/Jerteh-355",
     "kind": "model",
     "note": "Model with 5 likes on the Hub."
    },
    {
     "name": "jerteh/SrpELTeC",
     "url": "https://huggingface.co/datasets/jerteh/SrpELTeC",
     "kind": "dataset",
     "note": "Dataset with 3 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "gpt2-orao",
     "url": "https://huggingface.co/jerteh/gpt2-orao",
     "note": "Serbian GPT-2 model"
    },
    {
     "name": "Jerteh-355",
     "url": "https://huggingface.co/jerteh/Jerteh-355",
     "note": "Serbian language model"
    },
    {
     "name": "SrpELTeC",
     "url": "https://huggingface.co/datasets/jerteh/SrpELTeC",
     "note": "Serbian literary corpus"
    },
    {
     "name": "SrpKorNews",
     "url": "https://huggingface.co/datasets/jerteh/SrpKorNews",
     "note": "Serbian news corpus"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Mihailo Škorić",
     "role": "Model developer",
     "url": "https://huggingface.co/procesaur",
     "note": null
    },
    {
     "name": "Ranka Stanković",
     "role": "Member",
     "url": "https://huggingface.co/rankas",
     "note": null
    },
    {
     "name": "Milica Ikonić Nešić",
     "role": "Member",
     "url": "https://huggingface.co/MilicaIK",
     "note": null
    },
    {
     "name": "Jelena Graovac",
     "role": "Member",
     "url": "https://huggingface.co/jgraovac",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the society website.",
   "sources": [
    "https://jerteh.rs",
    "https://huggingface.co/jerteh",
    "https://github.com/jerteh"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "jerteh",
   "logo": true
  },
  {
   "orig_id": "kblab",
   "name": "KBLab",
   "alias": "National Library of Sweden",
   "full_name": "KBLab (National Library of Sweden)",
   "type": "institute",
   "founded": null,
   "country": "Sweden",
   "city": "Stockholm",
   "lat": 59.3293,
   "lon": 18.0686,
   "website": "https://www.kb.se/in-english/research-collaboration/kblab.html",
   "huggingface": "https://huggingface.co/KBLab",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Data-science lab within the National Library of Sweden that builds language and speech models from the library's collections. Its models are openly available on Hugging Face.",
   "tagline": "Data lab inside the National Library of Sweden that trains open Swedish language and speech models, including KB-BERT and KB-Whisper, on the library's collections.",
   "about": [
    "KBLab is a research department within the National Library of Sweden, set up in 2019 as a national research infrastructure for digital humanities and social science. It gives researchers access to the library's collections in structured form. The lab is directed by Love Börjeson.",
    "Using the legal-deposit and digitized holdings, KBLab trains Swedish models. Its 2020 paper by Martin Malmsten, Love Börjeson and Chris Haffenden introduced KB-BERT. The lab later released KB-Whisper, speech recognition models trained on more than 50,000 hours of Swedish speech, which it reports outperform OpenAI's Whisper at each size on Swedish.",
    "Models are published on Hugging Face, with the large KB-Whisper model under the Apache 2.0 license."
   ],
   "mission": "KBLab aims to make the National Library's collections usable for research through language and speech models and open tools, supporting digital humanities and social science.",
   "notice": null,
   "known_for": [
    "KB-BERT Swedish language model",
    "KB-Whisper Swedish speech recognition",
    "Models trained on national library collections"
   ],
   "focus": [
    "Swedish",
    "speech",
    "language models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "KBLab established",
     "note": "The National Library set up an integrated data lab for research.",
     "year": "2019"
    },
    {
     "title": "KB-BERT",
     "note": "Published the paper on building a Swedish BERT model from the library's collections.",
     "year": "2020"
    },
    {
     "title": "KB-Whisper",
     "note": "Released Swedish Whisper models trained on over 50,000 hours of speech, in sizes from tiny to large.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "KBLab/wav2vec2-large-voxrex-swedish",
     "url": "https://huggingface.co/KBLab/wav2vec2-large-voxrex-swedish",
     "kind": "model",
     "note": "Model: 728,037 downloads, 26 likes."
    },
    {
     "name": "KBLab/sentence-bert-swedish-cased",
     "url": "https://huggingface.co/KBLab/sentence-bert-swedish-cased",
     "kind": "model",
     "note": "Model: 38,229 downloads, 37 likes."
    },
    {
     "name": "KBLab/rixvox-v2",
     "url": "https://huggingface.co/datasets/KBLab/rixvox-v2",
     "kind": "dataset",
     "note": "Dataset: 5,545 downloads, 14 likes."
    },
    {
     "name": "KBLab/easytranscriber-demo",
     "url": "https://huggingface.co/spaces/KBLab/easytranscriber-demo",
     "kind": "space",
     "note": "Space: 4 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Playing with Words at the National Library of Sweden -- Making a Swedish BERT",
     "url": "https://arxiv.org/abs/2007.01658",
     "year": "2020"
    },
    {
     "title": "Hearing voices at the National Library -- a speech corpus and acoustic model for the Swedish language",
     "url": "https://arxiv.org/abs/2205.03026",
     "year": "2022"
    },
    {
     "title": "Swedish Whispers; Leveraging a Massive Speech Corpus for Swedish Speech Recognition",
     "url": "https://arxiv.org/abs/2505.17538",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "KB-Whisper large",
     "url": "https://huggingface.co/KBLab/kb-whisper-large",
     "note": "Swedish speech recognition model"
    },
    {
     "name": "Swedish sentence-BERT",
     "url": "https://huggingface.co/KBLab/sentence-bert-swedish-cased",
     "note": "Sentence embedding model"
    },
    {
     "name": "VoxRex wav2vec2",
     "url": "https://huggingface.co/KBLab/wav2vec2-large-voxrex-swedish",
     "note": "Swedish wav2vec2 model"
    },
    {
     "name": "RixVox",
     "url": "https://huggingface.co/datasets/KBLab/rixvox",
     "note": "Speech dataset from the Swedish Parliament"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face; check each card for licence terms.",
   "key_people": [
    {
     "name": "Love Börjeson",
     "role": "Director of KBLab",
     "url": "https://arxiv.org/abs/2007.01658",
     "note": null
    },
    {
     "name": "Martin Malmsten",
     "role": "Author of the KB-BERT paper",
     "url": "https://arxiv.org/abs/2007.01658",
     "note": null
    },
    {
     "name": "Chris Haffenden",
     "role": "Author of the KB-BERT paper",
     "url": "https://arxiv.org/abs/2007.01658",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the KBLab page on the National Library website.",
   "sources": [
    "https://www.kb.se/in-english/research-collaboration/kblab.html",
    "https://huggingface.co/KBLab"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "kblab",
   "logo": true
  },
  {
   "orig_id": "kinit",
   "name": "Kempelen Institute of Intelligent Technologies",
   "alias": "KInIT",
   "full_name": "Kempelen Institute of Intelligent Technologies (KInIT)",
   "type": "institute",
   "founded": 2020,
   "country": "Slovakia",
   "city": "Bratislava",
   "lat": 48.1486,
   "lon": 17.1077,
   "website": "https://www.kinit.sk",
   "huggingface": "https://huggingface.co/kinit",
   "github": "https://github.com/kinit-sk",
   "x": "https://x.com/KInIT_sk",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/kempelen-institute-of-intelligent-technologies",
   "contact_email": null,
   "summary": "Slovak independent research institute for AI that publishes Slovak language models, speech recognition models and benchmarks.",
   "tagline": "Independent Slovak non-profit research institute for intelligent technologies, known for SlovakBERT and open Slovak NLP benchmarks.",
   "about": [
    "The Kempelen Institute of Intelligent Technologies (KInIT) is an independent, non-profit institute in Bratislava. It was set up with backing from Slovak companies including ESET, Tatra banka, Innovatrics and Softec, and aims to connect the private and public sectors with multidisciplinary research. Mária Bieliková is Director General.",
    "Its research areas include web and user data processing and disinformation, natural language processing, data analysis for green energy, information security, and ethics of intelligent technologies. In 2021 its NLP team consulted on and evaluated SlovakBERT, the first public large neural language model for Slovak, trained with partner Gerulata Technologies.",
    "The institute publishes Slovak models and benchmarks, including the 2025 skLEP benchmark, on Hugging Face and GitHub."
   ],
   "mission": "KInIT states that it aims to improve Slovakia's competitiveness through multidisciplinary research on intelligent technologies, connect companies and public bodies, and encourage responsible innovation and talent development.",
   "notice": null,
   "known_for": [
    "SlovakBERT evaluation and release",
    "Slovak NLP benchmarks such as skLEP",
    "Disinformation and trustworthy AI research"
   ],
   "focus": [
    "Slovak",
    "language models",
    "speech",
    "misinformation"
   ],
   "contributions": [
    {
     "title": "SlovakBERT",
     "note": "Published the first public large Slovak neural language model with Gerulata Technologies, evaluated by KInIT.",
     "year": "2021"
    },
    {
     "title": "skLEP benchmark",
     "note": "Co-authored a Slovak language understanding benchmark.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "IMGTB",
     "url": "https://github.com/kinit-sk/IMGTB",
     "note": "Framework extending MGTBench for machine-generated text detection benchmarking (22 stars)"
    },
    {
     "name": "yaudit-recsys-2021",
     "url": "https://github.com/kinit-sk/yaudit-recsys-2021",
     "note": "Supplementary material for the YouTube audit paper at RecSys 2021 (20 stars)"
    },
    {
     "name": "medical-misinformation-dataset",
     "url": "https://github.com/kinit-sk/medical-misinformation-dataset",
     "note": "Medical misinformation dataset (16 stars)"
    },
    {
     "name": "mgt-detection-benchmark",
     "url": "https://github.com/kinit-sk/mgt-detection-benchmark",
     "note": "Multilingual machine-generated text detection benchmark (13 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "kinit/whisper-large-v3-turbo-sk",
     "url": "https://huggingface.co/kinit/whisper-large-v3-turbo-sk",
     "kind": "model",
     "note": "Slovak Whisper large-v3-turbo fine-tune."
    },
    {
     "name": "kinit/whisper-large-v3-sk",
     "url": "https://huggingface.co/kinit/whisper-large-v3-sk",
     "kind": "model",
     "note": "Slovak Whisper large-v3 fine-tune."
    },
    {
     "name": "kinit/gest",
     "url": "https://huggingface.co/datasets/kinit/gest",
     "kind": "dataset",
     "note": "Dataset published by KInIT."
    },
    {
     "name": "kinit/slovak-asr-leaderboard",
     "url": "https://huggingface.co/spaces/kinit/slovak-asr-leaderboard",
     "kind": "space",
     "note": "Leaderboard for Slovak speech recognition."
    }
   ],
   "papers": [
    {
     "title": "SlovakBERT: Slovak Masked Language Model",
     "url": "https://arxiv.org/abs/2109.15254",
     "year": "2021"
    },
    {
     "title": "IMGTB: A Framework for Machine-Generated Text Detection Benchmarking",
     "url": "https://aclanthology.org/2024.acl-demos.17/",
     "year": "2024"
    },
    {
     "title": "Auditing YouTube's Recommendation Algorithm for Misinformation Filter Bubbles",
     "url": "https://arxiv.org/abs/2210.10085",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "whisper-large-v3-turbo-sk",
     "url": "https://huggingface.co/kinit/whisper-large-v3-turbo-sk",
     "note": "Slovak ASR model"
    },
    {
     "name": "SlovakBERT sentiment",
     "url": "https://huggingface.co/kinit/slovakbert-sentiment-twitter",
     "note": "Slovak sentiment model"
    },
    {
     "name": "IMGTB",
     "url": "https://github.com/kinit-sk/IMGTB",
     "note": "Machine-generated text detection benchmark"
    }
   ],
   "openness": "Models, datasets and code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Mária Bieliková",
     "role": "Director General",
     "url": "https://www.kinit.sk/about-us/",
     "note": null
    },
    {
     "name": "Marián Šimko",
     "role": "Deputy Director General",
     "url": "https://www.kinit.sk/about-us/",
     "note": null
    },
    {
     "name": "Michal Kompan",
     "role": "Chief Research and Innovations Officer",
     "url": "https://www.kinit.sk/about-us/",
     "note": null
    },
    {
     "name": "Ivan Sokáč",
     "role": "Chief Operating and Strategy Officer",
     "url": "https://www.kinit.sk/about-us/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The institute's website could not be fetched during checking; use its GitHub and Hugging Face organisations for project contact.",
   "sources": [
    "https://www.kinit.sk",
    "https://huggingface.co/kinit",
    "https://github.com/kinit-sk",
    "https://arxiv.org/abs/2109.15254"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "kempelen-institute-of-intelligent-technologies",
   "logo": true
  },
  {
   "orig_id": "kotoba-technologies",
   "name": "Kotoba Technologies",
   "alias": null,
   "full_name": "Kotoba Technologies",
   "type": "startup",
   "founded": 2023,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://kotoba.tech",
   "huggingface": "https://huggingface.co/kotoba-tech",
   "github": "https://github.com/kotoba-tech",
   "x": "https://twitter.com/kotoba_tech",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Voice AI company, incorporated in 2023 with offices in San Francisco and Tokyo, that released the Kotoba-Whisper Japanese speech recognition models and builds real-time speech and translation models.",
   "tagline": "Voice AI company that released the Kotoba-Whisper Japanese speech recognition models and builds real-time speech and translation models.",
   "about": [
    "Kotoba Technologies was founded in 2023 by Noriyuki Kojima, chief executive, and Jungo Kasai, chief technology officer. The company is incorporated in San Francisco with a Tokyo operations office and describes itself as a frontier voice AI company with expertise in East Asian languages.",
    "It is known for Kotoba-Whisper, distilled Whisper models for Japanese speech recognition developed with Asahi Ushio. Version 1.0 appeared in April 2024, trained on the ReazonSpeech corpus, with version 2.0 following in September 2024. Its newer Koto model targets low-latency text-to-speech, speech-to-text and speech translation that can run on-premise or on-device.",
    "A Seed round of about 11.8 million dollars led by Globis Capital Partners was reported in 2025."
   ],
   "mission": "The company aims to build real-time speech models and simultaneous translation across languages, with a focus on East Asian languages.",
   "notice": "The company's homepage carries a notice dated 16 November 2026 saying that Kotoba Technologies is joining ElevenLabs.",
   "known_for": [
    "Kotoba-Whisper Japanese ASR models",
    "Real-time speech translation",
    "Koto voice model"
   ],
   "focus": [
    "speech",
    "Japanese"
   ],
   "contributions": [
    {
     "title": "Company incorporated",
     "note": "Kotoba Technologies, Inc. was incorporated in July and its Japanese entity in October.",
     "year": "2023"
    },
    {
     "title": "Kotoba-Whisper v1.0",
     "note": "Released distilled Whisper models for Japanese ASR.",
     "year": "2024"
    },
    {
     "title": "Kotoba-Whisper v2.0",
     "note": "Released an updated Japanese ASR model in September.",
     "year": "2024"
    },
    {
     "title": "Second seed round",
     "note": "Reported a seed round of about 11.8 million dollars led by Globis Capital Partners.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "kotoba-whisper",
     "url": "https://github.com/kotoba-tech/kotoba-whisper",
     "note": "Kotoba-Whisper code."
    },
    {
     "name": "kotomamba",
     "url": "https://github.com/kotoba-tech/kotomamba",
     "note": "Mamba training library."
    },
    {
     "name": "kotoba-speech-release",
     "url": "https://github.com/kotoba-tech/kotoba-speech-release",
     "note": "Speech model release."
    },
    {
     "name": "kotoba-recipes",
     "url": "https://github.com/kotoba-tech/kotoba-recipes",
     "note": "Continual pre-training recipes."
    }
   ],
   "hf_highlights": [
    {
     "name": "kotoba-whisper-v2.0",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0",
     "kind": "model",
     "note": "Main release."
    },
    {
     "name": "kotoba-whisper-v2.2",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v2.2",
     "kind": "model",
     "note": "Latest version."
    },
    {
     "name": "kotoba-whisper-v2.0-faster",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0-faster",
     "kind": "model",
     "note": "faster-whisper version."
    },
    {
     "name": "kotoba-whisper-eval",
     "url": "https://huggingface.co/datasets/kotoba-tech/kotoba-whisper-eval",
     "kind": "dataset",
     "note": "Evaluation data."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "kotoba-whisper-v2.2",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v2.2",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "kotoba-whisper-v2.0",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face, code on GitHub.",
   "key_people": [
    {
     "name": "Noriyuki Kojima",
     "role": "Co-founder and CEO",
     "url": "https://site.kotoba.tech/company",
     "note": null
    },
    {
     "name": "Jungo Kasai",
     "role": "Co-founder and CTO",
     "url": "https://site.kotoba.tech/company",
     "note": null
    },
    {
     "name": "Asahi Ushio",
     "role": "Developer of Kotoba-Whisper",
     "url": "https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website or GitHub.",
   "sources": [
    "https://huggingface.co/kotoba-tech",
    "https://kotoba.tech"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "kotoba-technologies",
   "mlat": 35.61876,
   "mlon": 139.69,
   "logo": true
  },
  {
   "orig_id": "kyutai",
   "name": "Kyutai",
   "alias": null,
   "full_name": "Kyutai",
   "type": "nonprofit",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://kyutai.org",
   "huggingface": "https://huggingface.co/kyutai",
   "github": "https://github.com/kyutai-labs",
   "x": "https://x.com/kyutai_labs",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/kyutai-labs",
   "contact_email": "contact@kyutai.org",
   "summary": "Paris-based open-science AI lab focused on speech and multimodal models. It publishes weights, code and papers for its work.",
   "tagline": "Paris non-profit open-science lab building speech-native and multimodal models, creator of Moshi, Mimi, Unmute and Helium.",
   "about": [
    "Kyutai is an open-science AI research lab in Paris, funded by the Iliad Group, the CMA CGM Group and Schmidt Sciences. Its team is led by Patrick Pérez as chief executive, with Alexandre Défossez as head of research and Laurent Mazaré as head of engineering. Edouard Grave and Hervé Jegou are listed among its alumni.",
    "The lab is known for Moshi, described as the first speech-native dialogue system, which processes speech directly rather than converting it to text and back. It also released Unmute, which gives any text model real-time listening and speaking, and Helium, a multilingual two-billion-parameter language model. Papers, code and weights are published openly.",
    "Its recent research output covers text-to-speech, video autoencoders and world models."
   ],
   "mission": "Kyutai aims to advance open research on speech and multimodal AI, publishing models, code and papers for the community.",
   "notice": null,
   "known_for": [
    "Moshi speech-native dialogue model",
    "Unmute streaming speech interface",
    "Helium multilingual language model"
   ],
   "focus": [
    "speech",
    "multimodal",
    "language models",
    "open science"
   ],
   "contributions": [
    {
     "title": "Lab launched",
     "note": "Kyutai was announced in Paris as a non-profit open-science lab.",
     "year": "2023"
    },
    {
     "title": "Moshi",
     "note": "Released a real-time speech-native dialogue model with open weights.",
     "year": "2024"
    },
    {
     "title": "Mimi codec",
     "note": "Published the streaming neural audio codec that underpins Moshi.",
     "year": "2024"
    },
    {
     "title": "Unmute",
     "note": "Released a system adding real-time speech input and output to text models.",
     "year": "2025"
    },
    {
     "title": "Helium 1",
     "note": "Released a modular multilingual language model of 2B parameters.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "moshi",
     "url": "https://github.com/kyutai-labs/moshi",
     "note": "Moshi is a speech-text foundation model and full-duplex spoken dialogue framework. It uses Mimi, a state-of-th (11,542 stars)"
    },
    {
     "name": "pocket-tts",
     "url": "https://github.com/kyutai-labs/pocket-tts",
     "note": "A TTS that fits in your CPU (and pocket) (10,320 stars)"
    },
    {
     "name": "delayed-streams-modeling",
     "url": "https://github.com/kyutai-labs/delayed-streams-modeling",
     "note": "Kyutai's Speech-To-Text and Text-To-Speech models based on the Delayed Streams Modeling framework. (3,053 stars)"
    },
    {
     "name": "unmute",
     "url": "https://github.com/kyutai-labs/unmute",
     "note": "Make text LLMs listen and speak (1,553 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "kyutai/moshiko-pytorch-bf16",
     "url": "https://huggingface.co/kyutai/moshiko-pytorch-bf16",
     "kind": "model",
     "note": "Model: 308,499 downloads, 266 likes."
    },
    {
     "name": "kyutai/tts-1.6b-en_fr",
     "url": "https://huggingface.co/kyutai/tts-1.6b-en_fr",
     "kind": "model",
     "note": "Model: 126,722 downloads, 384 likes."
    },
    {
     "name": "kyutai/DailyTalkContiguous",
     "url": "https://huggingface.co/datasets/kyutai/DailyTalkContiguous",
     "kind": "dataset",
     "note": "Dataset: 2,646 downloads, 22 likes."
    },
    {
     "name": "kyutai/hibiki-samples",
     "url": "https://huggingface.co/spaces/kyutai/hibiki-samples",
     "kind": "space",
     "note": "Space: 56 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Moshi: a speech-text foundation model for real-time dialogue",
     "url": "https://arxiv.org/abs/2410.00037",
     "year": "2024"
    },
    {
     "title": "High-Fidelity Simultaneous Speech-To-Speech Translation",
     "url": "https://arxiv.org/abs/2502.03382",
     "year": "2025"
    },
    {
     "title": "Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling",
     "url": "https://arxiv.org/abs/2509.08753",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Moshi",
     "url": "https://github.com/kyutai-labs/moshi",
     "note": "Full-duplex speech-text dialogue model and framework"
    },
    {
     "name": "Mimi codec",
     "url": "https://huggingface.co/kyutai/mimi",
     "note": "Streaming neural audio codec"
    },
    {
     "name": "Pocket TTS",
     "url": "https://github.com/kyutai-labs/pocket-tts",
     "note": "Text-to-speech model that runs on a CPU"
    },
    {
     "name": "Delayed Streams Modeling STT/TTS",
     "url": "https://github.com/kyutai-labs/delayed-streams-modeling",
     "note": "Speech-to-text and text-to-speech models"
    }
   ],
   "openness": "Models and code are released openly, mostly under Apache-2.0 or MIT licences, with technical papers.",
   "key_people": [
    {
     "name": "Patrick Pérez",
     "role": "Chief Executive Officer",
     "url": "https://ptrckprz.github.io/",
     "note": null
    },
    {
     "name": "Alexandre Défossez",
     "role": "Head of Research",
     "url": "https://scholar.google.fr/citations?user=DubNUU0AAAAJ",
     "note": null
    },
    {
     "name": "Laurent Mazaré",
     "role": "Head of Engineering",
     "url": "https://github.com/LaurentMazare",
     "note": null
    },
    {
     "name": "Neil Zeghidour",
     "role": "Audio Research Advisor",
     "url": "https://scholar.google.com/citations?user=fiJamZ0AAAAJ",
     "note": null
    },
    {
     "name": "Václav Volhejn",
     "role": "Technical staff",
     "url": "https://vvolhejn.com/",
     "note": null
    },
    {
     "name": "Amélie Royer",
     "role": "Technical staff",
     "url": "https://ameroyer.github.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact address printed on the website.",
   "sources": [
    "https://kyutai.org",
    "https://huggingface.co/kyutai",
    "https://github.com/kyutai-labs",
    "https://arxiv.org/abs/2410.00037"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "kyutai",
   "mlat": 48.94424,
   "mlon": 2.43467,
   "logo": true
  },
  {
   "orig_id": "laion",
   "name": "LAION",
   "alias": "incl. OpenAssistant",
   "full_name": "LAION (incl. OpenAssistant)",
   "type": "nonprofit",
   "founded": 2021,
   "country": "Germany",
   "city": "Hamburg",
   "lat": 53.5511,
   "lon": 9.9937,
   "website": "https://laion.ai",
   "huggingface": "https://huggingface.co/laion",
   "github": "https://github.com/LAION-AI",
   "x": null,
   "discord": "https://discord.com/invite/eq3cAMZtCC",
   "linkedin": null,
   "contact_email": "contact@laion.ai",
   "summary": "LAION is a non-profit making datasets, tools and models openly available for research. It also co-organised the OpenAssistant crowd-sourced conversation project.",
   "tagline": "German non-profit that releases large open datasets, tools and models for AI research, best known for LAION-400M, LAION-5B and OpenCLIP.",
   "about": [
    "LAION, the Large-scale Artificial Intelligence Open Network, is a non-profit research network based in Germany that makes machine learning resources available to the public. It is supported by donations and public research grants. Its volunteers include Christoph Schuhmann, Jenia Jitsev and Romain Beaumont.",
    "In 2021 it released LAION-400M, an image-text dataset recreating the kind of data used for CLIP, followed by LAION-5B in 2022, the largest openly available image-text dataset at the time. Stable Diffusion was trained on it, and LAION members helped train OpenCLIP models on it. In 2023 it co-organised the OpenAssistant crowd-sourced conversation project.",
    "After Stanford researchers reported suspected abuse material links in LAION-5B in December 2023, LAION withdrew its datasets and in August 2024 released a cleaned Re-LAION-5B."
   ],
   "mission": "LAION states that it makes large-scale machine learning resources, including datasets, tools and open models, available to the public for research and education.",
   "notice": null,
   "known_for": [
    "LAION-5B image-text dataset",
    "OpenCLIP models",
    "OpenAssistant"
   ],
   "focus": [
    "datasets",
    "vision",
    "speech",
    "language models"
   ],
   "contributions": [
    {
     "title": "LAION-400M",
     "note": "Released 400 million image-text pairs filtered with CLIP.",
     "year": "2021"
    },
    {
     "title": "LAION-5B",
     "note": "Released over 5 billion image-text pairs, used to train Stable Diffusion.",
     "year": "2022"
    },
    {
     "title": "Reproducible scaling laws",
     "note": "Published OpenCLIP scaling-law research using LAION-2B.",
     "year": "2022"
    },
    {
     "title": "OpenAssistant",
     "note": "Released an open chat assistant and dataset built by volunteers.",
     "year": "2023"
    },
    {
     "title": "Dataset takedown",
     "note": "Removed LAION-5B and LAION-400M temporarily after a report of illegal content.",
     "year": "2023"
    },
    {
     "title": "Re-LAION-5B",
     "note": "Released a cleaned version of the dataset.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "Open-Assistant",
     "url": "https://github.com/LAION-AI/Open-Assistant",
     "note": "Chat-based assistant project; most-starred repo."
    },
    {
     "name": "CLAP",
     "url": "https://github.com/LAION-AI/CLAP",
     "note": "Contrastive Language-Audio Pretraining."
    },
    {
     "name": "CLIP_benchmark",
     "url": "https://github.com/LAION-AI/CLIP_benchmark",
     "note": "Evaluation for CLIP-like models."
    },
    {
     "name": "aesthetic-predictor",
     "url": "https://github.com/LAION-AI/aesthetic-predictor",
     "note": "Linear aesthetic estimator on CLIP embeddings."
    }
   ],
   "hf_highlights": [
    {
     "name": "clap-htsat-fused",
     "url": "https://huggingface.co/laion/clap-htsat-fused",
     "kind": "model",
     "note": "CLAP audio-text model."
    },
    {
     "name": "CLIP-ViT-B-32-laion2B-s34B-b79K",
     "url": "https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K",
     "kind": "model",
     "note": "OpenCLIP ViT-B/32 trained on LAION-2B."
    },
    {
     "name": "CLIP-ViT-L-14-laion2B-s32B-b82K",
     "url": "https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K",
     "kind": "model",
     "note": "OpenCLIP ViT-L/14 trained on LAION-2B."
    }
   ],
   "papers": [
    {
     "title": "LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs",
     "url": "https://arxiv.org/abs/2111.02114",
     "year": "2021"
    },
    {
     "title": "LAION-5B: An open large-scale dataset for training next generation image-text models",
     "url": "https://arxiv.org/abs/2210.08402",
     "year": "2022"
    },
    {
     "title": "Reproducible scaling laws for contrastive language-image learning",
     "url": "https://arxiv.org/abs/2212.07143",
     "year": "2022"
    },
    {
     "title": "OpenFlamingo: An Open-Source Framework for Training Large Autoregressive Vision-Language Models",
     "url": "https://arxiv.org/abs/2308.01390",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "LAION-5B paper",
     "url": "https://arxiv.org/abs/2210.08402",
     "note": "Image-text dataset"
    },
    {
     "name": "CLAP",
     "url": "https://github.com/LAION-AI/CLAP",
     "note": "Audio-language model"
    },
    {
     "name": "OpenAssistant oasst1",
     "url": "https://huggingface.co/datasets/OpenAssistant/oasst1",
     "note": "Conversation dataset"
    },
    {
     "name": "CLIP models",
     "url": "https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K",
     "note": "OpenCLIP checkpoints"
    }
   ],
   "openness": "Datasets, models and code are open with mixed licences.",
   "key_people": [
    {
     "name": "Christoph Schuhmann",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/LAION",
     "note": null
    },
    {
     "name": "Jenia Jitsev",
     "role": "Co-founder, Jülich Supercomputing Centre",
     "url": "https://en.wikipedia.org/wiki/LAION",
     "note": null
    },
    {
     "name": "Romain Beaumont",
     "role": "Core contributor, OpenCLIP model training",
     "url": "https://arxiv.org/abs/2210.08402",
     "note": null
    },
    {
     "name": "Richard Vencu",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/LAION",
     "note": null
    },
    {
     "name": "Mehdi Cherti",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/LAION",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://laion.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://laion.ai",
    "https://huggingface.co/laion",
    "https://github.com/LAION-AI",
    "https://arxiv.org/abs/2304.07327"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "laion",
   "logo": true
  },
  {
   "orig_id": "lang-uk",
   "name": "lang-uk",
   "alias": null,
   "full_name": "lang-uk",
   "type": "collective",
   "founded": 2016,
   "country": "Ukraine",
   "city": "Kyiv",
   "lat": 50.4501,
   "lon": 30.5234,
   "website": "https://lang.org.ua",
   "huggingface": "https://huggingface.co/lang-uk",
   "github": "https://github.com/lang-uk",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Volunteer community for Ukrainian natural language processing that publishes open datasets, tools and models.",
   "tagline": "Volunteer community of Ukrainian NLP specialists building open corpora, models and tools for the Ukrainian language.",
   "about": [
    "lang-uk is an informal community of programmers, linguists and researchers that supports existing projects and starts new ones to collect Ukrainian text corpora and other language data. Its work is hosted in the lang-uk GitHub organization and on Hugging Face, and the community is open to new members. Ukrainian media first covered its results after its first year of work in 2015.",
    "Co-founder Vsevolod Dyomkin has described the project's main task as collecting and maintaining data, models and tools for processing Ukrainian. Mariana Romanyshyn led a group of linguists who began collecting open Ukrainian text data, and Dmytro Chaplynskyi is credited on the UberText 2.0 corpus paper. The Ukrainian Catholic University and the Kyiv School of Economics are listed as partners.",
    "Its resources serve scientists, entrepreneurs and journalists, covering services, models, libraries, dictionaries, gazetteers, corpora and datasets."
   ],
   "mission": "The community aims to create the conditions for full processing of Ukrainian-language text by gathering and openly sharing corpora, tools and models.",
   "notice": null,
   "known_for": [
    "Open Ukrainian corpora and datasets",
    "UberText 2.0 corpus",
    "Ukrainian NLP tooling and models"
   ],
   "focus": [
    "Ukrainian",
    "NLP",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "First-year report",
     "note": "Published an account of the community's first year of results on DOU.",
     "year": "2015"
    },
    {
     "title": "Project announcement",
     "note": "Announced the lang-uk initiative to build conditions for Ukrainian text processing.",
     "year": "2016"
    },
    {
     "title": "UberText 2.0",
     "note": "Contributors published the UberText 2.0 Ukrainian text corpus at the UNLP workshop.",
     "year": "2023"
    },
    {
     "title": "UNLP workshop",
     "note": "Members helped organize the Ukrainian NLP workshop, including its 2024 edition.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "ner-uk",
     "url": "https://github.com/lang-uk/ner-uk",
     "note": "Ukrainian NER annotation project (95 stars)"
    },
    {
     "name": "tokenize-uk",
     "url": "https://github.com/lang-uk/tokenize-uk",
     "note": "Tokenizer for Ukrainian sentences and words (71 stars)"
    },
    {
     "name": "ukrainian-word-stress",
     "url": "https://github.com/lang-uk/ukrainian-word-stress",
     "note": "Adds word stress to Ukrainian texts (66 stars)"
    },
    {
     "name": "tone-dict-uk",
     "url": "https://github.com/lang-uk/tone-dict-uk",
     "note": "Ukrainian tone dictionary (46 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "lang-uk/ukr-paraphrase-multilingual-mpnet-base",
     "url": "https://huggingface.co/lang-uk/ukr-paraphrase-multilingual-mpnet-base",
     "kind": "model",
     "note": "Ukrainian sentence embedding model."
    },
    {
     "name": "lang-uk/malyuk",
     "url": "https://huggingface.co/datasets/lang-uk/malyuk",
     "kind": "dataset",
     "note": "Combined Ukrainian text corpus."
    },
    {
     "name": "lang-uk/dragoman",
     "url": "https://huggingface.co/lang-uk/dragoman",
     "kind": "model",
     "note": "Ukrainian-English translation model."
    },
    {
     "name": "lang-uk/ukrainian-llm-leaderboard",
     "url": "https://huggingface.co/spaces/lang-uk/ukrainian-llm-leaderboard",
     "kind": "space",
     "note": "Ukrainian LLM leaderboard."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "ner-uk",
     "url": "https://github.com/lang-uk/ner-uk",
     "note": "Ukrainian NER annotation project"
    },
    {
     "name": "tokenize-uk",
     "url": "https://github.com/lang-uk/tokenize-uk",
     "note": "Ukrainian tokenizer"
    },
    {
     "name": "Dragoman",
     "url": "https://huggingface.co/lang-uk/dragoman",
     "note": "Ukrainian-English translation model"
    },
    {
     "name": "Ukrainian LLM Leaderboard",
     "url": "https://huggingface.co/spaces/lang-uk/ukrainian-llm-leaderboard",
     "note": "Leaderboard for Ukrainian LLMs"
    }
   ],
   "openness": "Code, models and datasets are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Vsevolod Dyomkin",
     "role": "Co-founder of lang-uk",
     "url": "https://lang.org.ua",
     "note": null
    },
    {
     "name": "Mariana Romanyshyn",
     "role": "Led early open data collection; UNLP organizer",
     "url": "https://dou.ua/lenta/columns/lang-uk/",
     "note": null
    },
    {
     "name": "Dmytro Chaplynskyi",
     "role": "Author of the UberText 2.0 paper, affiliated with lang-uk",
     "url": "https://aclanthology.org/2023.unlp-1.1.pdf",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The community's site has a contacts page and a manifesto; its work lives in the lang-uk GitHub organisation.",
   "sources": [
    "https://lang.org.ua",
    "https://huggingface.co/lang-uk",
    "https://github.com/lang-uk"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "lang-uk",
   "logo": true
  },
  {
   "orig_id": "lvl-reykjavik",
   "name": "Language and Voice Lab",
   "alias": "Reykjavik University",
   "full_name": "Language and Voice Lab (Reykjavik University)",
   "type": "academic",
   "founded": null,
   "country": "Iceland",
   "city": "Reykjavik",
   "lat": 64.1466,
   "lon": -21.9426,
   "website": "https://lvl.ru.is",
   "huggingface": "https://huggingface.co/language-and-voice-lab",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Reykjavik University lab building open Icelandic speech recognition and synthesis models and large speech corpora.",
   "tagline": "The Language and Voice Lab at Reykjavik University builds open Icelandic speech corpora and recognition and synthesis models.",
   "about": [
    "The Language and Voice Lab develops speech technology for Icelandic. It ran the crowd-sourced Samrómur corpus with Almannarómur, the Icelandic Center for Language Technology, and the latest release holds about 2,233 hours of recordings.",
    "It also released the Althingi parliamentary corpus of about 542 hours covering 2005 to 2016, the first public Icelandic corpus of children's voices, and synthetic data generated with its TTS systems. The datasets and models are licensed openly on Hugging Face, OpenSLR and LDC."
   ],
   "mission": "To make Icelandic speech technology openly available, with large corpora and models for recognition and synthesis.",
   "notice": null,
   "known_for": [
    "Samrómur Icelandic speech corpus",
    "Althingi parliamentary speech corpus",
    "Open Icelandic ASR and TTS models"
   ],
   "focus": [
    "Icelandic",
    "speech recognition",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Althingi corpus on Hugging Face",
     "note": "The 542-hour Althingi speech corpus was published on the hub.",
     "year": "2022"
    },
    {
     "title": "Raddrómur corpus",
     "note": "The Raddrómur Icelandic speech corpus was created at the lab.",
     "year": "2022"
    },
    {
     "title": "Samrómur Children",
     "note": "The first public Icelandic corpus with children's voices was released under Creative Commons.",
     "year": "2022"
    },
    {
     "title": "Whisper Icelandic models",
     "note": "Fine-tuned Whisper Large models for Icelandic ASR were released.",
     "year": "2023"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "language-and-voice-lab/whisper-large-icelandic-62640-steps-967h",
     "url": "https://huggingface.co/language-and-voice-lab/whisper-large-icelandic-62640-steps-967h",
     "kind": "model",
     "note": "Model with 5 likes on the Hub."
    },
    {
     "name": "language-and-voice-lab/wav2vec2-large-xlsr-53-icelandic-ep30-967h",
     "url": "https://huggingface.co/language-and-voice-lab/wav2vec2-large-xlsr-53-icelandic-ep30-967h",
     "kind": "model",
     "note": "Model with 3 likes on the Hub."
    },
    {
     "name": "language-and-voice-lab/samromur_children",
     "url": "https://huggingface.co/datasets/language-and-voice-lab/samromur_children",
     "kind": "dataset",
     "note": "Dataset with 9 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Whisper Icelandic",
     "url": "https://huggingface.co/language-and-voice-lab/whisper-large-icelandic-62640-steps-967h",
     "note": "Fine-tuned Whisper for Icelandic"
    },
    {
     "name": "wav2vec2 Icelandic",
     "url": "https://huggingface.co/language-and-voice-lab/wav2vec2-large-xlsr-53-icelandic-ep30-967h",
     "note": "Icelandic ASR model"
    },
    {
     "name": "Samromur children",
     "url": "https://huggingface.co/datasets/language-and-voice-lab/samromur_children",
     "note": "Crowdsourced Icelandic speech"
    },
    {
     "name": "Samromur Milljon",
     "url": "https://huggingface.co/datasets/language-and-voice-lab/samromur_milljon",
     "note": "Icelandic speech corpus"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face under the lab organisation.",
   "key_people": [
    {
     "name": "Jón Guðnason",
     "role": "Contact person for lab corpora",
     "url": "https://catalog.ldc.upenn.edu/docs/LDC2023S05/README.txt",
     "note": null
    },
    {
     "name": "Carlos Daniel Hernández Mena",
     "role": "Speech researcher",
     "url": "https://huggingface.co/carlosdanielhernandezmena",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the lab website.",
   "sources": [
    "https://lvl.ru.is",
    "https://huggingface.co/language-and-voice-lab"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "language-and-voice-lab",
   "mlat": 64.11124,
   "mlon": -21.9426,
   "logo": true
  },
  {
   "orig_id": "techiaith",
   "name": "Language Technologies Unit",
   "alias": "Bangor University",
   "full_name": "Language Technologies Unit (Bangor University)",
   "type": "academic",
   "founded": null,
   "country": "United Kingdom",
   "city": "Bangor",
   "lat": 53.2274,
   "lon": -4.1293,
   "website": "https://techiaith.cymru",
   "huggingface": "https://huggingface.co/techiaith",
   "github": "https://github.com/techiaith",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Welsh language technology unit at Bangor University releasing Welsh speech, translation and language models.",
   "tagline": "The Language Technologies Unit at Bangor University builds open Welsh speech, translation and language tools.",
   "about": [
    "Uned Technolegau Iaith is a self-funded research unit at Bangor University that develops digital language resources for Welsh, other Celtic languages and multilingual settings. Delyth Prys is head of the unit, and Gruffudd Prys, Dewi Bryn Jones and colleagues author its work.",
    "It runs the National Language Technologies Portal and maintains Welsh speech recognition, text-to-speech, translation and terminology tools. Its public repositories date back to 2014, and it has worked with Welsh Common Voice, the Welsh government and S4C."
   ],
   "mission": "To make language technology available for Welsh and to share the tools openly.",
   "notice": null,
   "known_for": [
    "Welsh speech recognition and synthesis",
    "Welsh machine translation",
    "Welsh National Language Technologies Portal"
   ],
   "focus": [
    "Welsh",
    "speech recognition",
    "machine translation"
   ],
   "contributions": [
    {
     "title": "GitHub organization opened",
     "note": "The unit began publishing code on GitHub.",
     "year": "2014"
    },
    {
     "title": "MosesSMT Welsh translation",
     "note": "Docker tooling for Welsh statistical machine translation was published.",
     "year": "2015"
    },
    {
     "title": "New resources for Welsh",
     "note": "The unit released resources for communicating with computers in Welsh, funded by the Welsh government and S4C.",
     "year": "2016"
    },
    {
     "title": "DeepSpeech Welsh recipes",
     "note": "Docker tooling for training Welsh speech recognition was released.",
     "year": "2018"
    },
    {
     "title": "HuggingFace Welsh STT",
     "note": "Tooling for Welsh speech recognition with Hugging Face was released.",
     "year": "2021"
    }
   ],
   "github_highlights": [
    {
     "name": "docker-moses-smt",
     "url": "https://github.com/techiaith/docker-moses-smt",
     "note": "Hwyluso cyfieithu peirianyddol MosesSMT i'r Gymraeg // Making MosesSMT machine translation easier for Welsh (a (16 stars)"
    },
    {
     "name": "docker-huggingface-stt-cy",
     "url": "https://github.com/techiaith/docker-huggingface-stt-cy",
     "note": "Adnabod lleferydd Cymraeg i'r Gymraeg gyda HuggingFace  // Speech Recognition for Welsh with HuggingFace  (13 stars)"
    },
    {
     "name": "pyfestival",
     "url": "https://github.com/techiaith/pyfestival",
     "note": "Amlapiwr Python C ar gyfer hwyluso rhaglennu gyda Festival | A Python C wrapper for simple coding with Festiva (11 stars)"
    },
    {
     "name": "lecsicon-cymraeg-bangor",
     "url": "https://github.com/techiaith/lecsicon-cymraeg-bangor",
     "note": "Lecsicon cynhwysfawr o eirffurfiau'r Gymraeg yn seiliedig ar ddata gwirydd sillafu a gramadeg Cysill // A comp (11 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "techiaith/mt-dspec-health-en-cy",
     "url": "https://huggingface.co/techiaith/mt-dspec-health-en-cy",
     "kind": "model",
     "note": "Model with 2 likes on the Hub."
    },
    {
     "name": "techiaith/fullstop-welsh-punctuation-prediction",
     "url": "https://huggingface.co/techiaith/fullstop-welsh-punctuation-prediction",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "techiaith/banc-trawsgrifiadau-bangor",
     "url": "https://huggingface.co/datasets/techiaith/banc-trawsgrifiadau-bangor",
     "kind": "dataset",
     "note": "Dataset with 4 likes on the Hub."
    },
    {
     "name": "techiaith/mimic",
     "url": "https://huggingface.co/spaces/techiaith/mimic",
     "kind": "space",
     "note": "Space with 0 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Welsh ASR",
     "url": "https://github.com/techiaith/docker-huggingface-stt-cy",
     "note": "Welsh speech recognition with Hugging Face"
    },
    {
     "name": "Llama-3 Welsh CPT",
     "url": "https://huggingface.co/techiaith/llama-3-8b-instruct-cpt-cy",
     "note": "Welsh continued pre-training model"
    },
    {
     "name": "Bangor transcripts",
     "url": "https://huggingface.co/datasets/techiaith/banc-trawsgrifiadau-bangor",
     "note": "Welsh speech transcripts"
    },
    {
     "name": "Macsen",
     "url": "https://github.com/techiaith/macsen",
     "note": "Welsh digital assistant prototype"
    }
   ],
   "openness": "Many repositories are public on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Delyth Prys",
     "role": "Head of Unit",
     "url": "https://techiaith.cymru",
     "note": null
    },
    {
     "name": "Gruffudd Prys",
     "role": "Researcher",
     "url": "https://huggingface.co/GruffPrys",
     "note": null
    },
    {
     "name": "Dewi Bryn Jones",
     "role": "Researcher",
     "url": "https://huggingface.co/DewiBrynJones",
     "note": null
    },
    {
     "name": "Stefano Ghazzali",
     "role": "Researcher",
     "url": "https://huggingface.co/stefanoghazzali",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the unit website.",
   "sources": [
    "https://techiaith.cymru",
    "https://huggingface.co/techiaith",
    "https://github.com/techiaith"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "language-technologies-unit",
   "logo": true
  },
  {
   "orig_id": "ltg-oslo",
   "name": "Language Technology Group, University of Oslo",
   "alias": null,
   "full_name": "Language Technology Group, University of Oslo",
   "type": "academic",
   "founded": null,
   "country": "Norway",
   "city": "Oslo",
   "lat": 59.9139,
   "lon": 10.7522,
   "website": "https://www.mn.uio.no/ifi/english/research/groups/ltg/",
   "huggingface": "https://huggingface.co/ltg",
   "github": "https://github.com/ltgoslo",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "NLP research group at the University of Oslo that has released Norwegian corpora, benchmarks and NorMistral language models.",
   "tagline": "University of Oslo research group in natural language processing behind the NorBERT models, the NorBench benchmark and Norwegian generative models.",
   "about": [
    "The Language Technology Group (LTG) is part of the Department of Informatics at the University of Oslo, led by Lilja Øvrelid. Its members include Stephan Oepen, Erik Velldal, Andrey Kutuzov, Yves Scherrer, Pierre Lison and David Samuel. The group works on natural language processing and large language models, including for Norwegian.",
    "It released NorBERT in January 2021 and NorBERT 2 in February 2022, then the NorBERT 3 family and the NorBench benchmark in 2023. The group has also trained generative Norwegian models such as NorMistral and NorBLOOM. Projects it participates in include HPLT, OpenEuroLLM, NorwAI, Integreat and MediaFutures.",
    "The group coordinates the Nordic Language Processing Laboratory and publishes code and models on GitHub and Hugging Face."
   ],
   "mission": "LTG develops language technology and language models, with a focus on Norwegian and openly available resources for evaluation and training.",
   "notice": null,
   "known_for": [
    "NorBERT language models",
    "NorBench benchmark",
    "NorMistral and NorBLOOM models"
   ],
   "focus": [
    "Norwegian",
    "language models",
    "benchmarks",
    "datasets"
   ],
   "contributions": [
    {
     "title": "NorBERT",
     "note": "Released the first large-scale Norwegian contextualised language model, described at NoDaLiDa 2021.",
     "year": "2021"
    },
    {
     "title": "NorBERT 2",
     "note": "Released an updated model trained on about 15 billion words of C4 and NCC.",
     "year": "2022"
    },
    {
     "title": "NorBERT 3 and NorBench",
     "note": "Released a family of four model sizes together with a benchmark for Norwegian.",
     "year": "2023"
    },
    {
     "title": "NorMistral and NorBLOOM",
     "note": "Trained generative Norwegian language models.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "norne",
     "url": "https://github.com/ltgoslo/norne",
     "note": "Norwegian Named Entities annotations on top of NDT (Norwegian Dependency Treebank) (72 stars)"
    },
    {
     "name": "gpt-bert",
     "url": "https://github.com/ltgoslo/gpt-bert",
     "note": "Official implementation of \"GPT or BERT: why not both?\" (64 stars)"
    },
    {
     "name": "norec",
     "url": "https://github.com/ltgoslo/norec",
     "note": "Norwegian Review Corpus (51 stars)"
    },
    {
     "name": "simple_elmo",
     "url": "https://github.com/ltgoslo/simple_elmo",
     "note": "Simple library to work with pre-trained ELMo models in TensorFlow (51 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "ltg/normistral-7b-warm",
     "url": "https://huggingface.co/ltg/normistral-7b-warm",
     "kind": "model",
     "note": "Model: 5,305 downloads, 31 likes."
    },
    {
     "name": "ltg/gpt-bert-babylm-base",
     "url": "https://huggingface.co/ltg/gpt-bert-babylm-base",
     "kind": "model",
     "note": "Model: 5,257 downloads, 10 likes."
    },
    {
     "name": "ltg/nrk_quiz_qa",
     "url": "https://huggingface.co/datasets/ltg/nrk_quiz_qa",
     "kind": "dataset",
     "note": "Dataset: 9,391 downloads, 3 likes."
    },
    {
     "name": "ltg/nor-ud",
     "url": "https://huggingface.co/spaces/ltg/nor-ud",
     "kind": "space",
     "note": "Space: 2 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "NorBench -- A Benchmark for Norwegian Language Models",
     "url": "https://arxiv.org/abs/2305.03880",
     "year": "2023"
    },
    {
     "title": "GPT or BERT: why not both?",
     "url": "https://arxiv.org/abs/2410.24159",
     "year": "2024"
    },
    {
     "title": "Small Languages, Big Models: A Study of Continual Training on Languages of Norway",
     "url": "https://arxiv.org/abs/2412.06484",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "NorMistral-7B",
     "url": "https://huggingface.co/ltg/normistral-7b-warm",
     "note": "Norwegian continued-pretraining model"
    },
    {
     "name": "NorNE",
     "url": "https://github.com/ltgoslo/norne",
     "note": "Norwegian named entity annotations"
    },
    {
     "name": "GPT-BERT",
     "url": "https://github.com/ltgoslo/gpt-bert",
     "note": "Hybrid GPT/BERT training code"
    }
   ],
   "openness": "Models, corpora and code are public on Hugging Face and GitHub under mostly open licences.",
   "key_people": [
    {
     "name": "Lilja Øvrelid",
     "role": "Group leader",
     "url": "https://www.mn.uio.no/ifi/english/research/groups/ltg/",
     "note": null
    },
    {
     "name": "Stephan Oepen",
     "role": "Professor",
     "url": "https://www.mn.uio.no/ifi/english/research/groups/ltg/",
     "note": null
    },
    {
     "name": "Erik Velldal",
     "role": "Professor",
     "url": "https://www.mn.uio.no/ifi/english/research/groups/ltg/",
     "note": null
    },
    {
     "name": "Andrey Kutuzov",
     "role": "Researcher, NorBERT author",
     "url": "https://arxiv.org/pdf/2104.06546.pdf",
     "note": null
    },
    {
     "name": "David Samuel",
     "role": "Researcher, lead author of NorBench",
     "url": "https://arxiv.org/pdf/2305.03880",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://www.mn.uio.no/ifi/english/research/groups/ltg/",
    "https://huggingface.co/ltg",
    "https://github.com/ltgoslo",
    "https://arxiv.org/abs/2305.03880"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "language-technology-group-university-of-oslo",
   "logo": true
  },
  {
   "orig_id": "latam-gpt",
   "name": "Latam-GPT",
   "alias": "CENIA",
   "full_name": "Latam-GPT (CENIA)",
   "type": "institute",
   "founded": 2024,
   "country": "Chile",
   "city": "Santiago",
   "lat": -33.4489,
   "lon": -70.6693,
   "website": "https://latamgpt.org",
   "huggingface": "https://huggingface.co/latam-gpt",
   "github": "https://github.com/latam-gpt",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "latam-gpt@cenia.cl",
   "summary": "Latam-GPT is a Latin American language model project coordinated by CENIA, Chile's national AI centre. It publishes models, a regional text corpus, benchmarks and data-curation tools openly.",
   "tagline": "Latin American open language model project coordinated by Chile's national AI centre CENIA, released in 2026 with a regional text corpus.",
   "about": [
    "Latam-GPT is a regional initiative to build a language model entirely within Latin America, for Latin America. It is coordinated by CENIA, the Chilean national AI centre, whose director Álvaro Soto spoke for the project at its launch. Reports describe a network of more than thirty institutions across the region.",
    "Chile launched the model on 10 February 2026 as open infrastructure for developers, trained on more than eight terabytes of Latin American data in Spanish and Portuguese, with indigenous languages planned. Press reports give its cost as about 550,000 dollars, funded by CAF and CENIA, with a first version trained on cloud infrastructure and later versions planned for a supercomputer at the University of Tarapacá.",
    "Models, data and tools are published on GitHub and Hugging Face, and the model is free for companies and public institutions to build on."
   ],
   "mission": "The project aims to give the region a model that represents its languages, culture and context, and to let developers build region-specific applications without depending on models made elsewhere.",
   "notice": null,
   "known_for": [
    "First open Latin American LLM",
    "Regional Spanish and Portuguese corpus",
    "Led by CENIA in Chile"
   ],
   "focus": [
    "language models",
    "datasets",
    "evaluation",
    "low-resource languages"
   ],
   "contributions": [
    {
     "title": "Latam-GPT launch",
     "note": "Chile launched the open Latam-GPT model on 10 February.",
     "year": "2026"
    },
    {
     "title": "Regional corpus",
     "note": "Assembled more than eight terabytes of Latin American text data for training.",
     "year": "2026"
    },
    {
     "title": "Regional funding model",
     "note": "Financed training through CAF and CENIA at a reported cost of 550,000 dollars.",
     "year": "2026"
    },
    {
     "title": "Institutional network",
     "note": "Brought together over thirty institutions from across the region.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "llm-data-eval",
     "url": "https://github.com/latam-gpt/llm-data-eval",
     "note": "LLM-aided filtering."
    },
    {
     "name": "toxicity-classifier",
     "url": "https://github.com/latam-gpt/toxicity-classifier",
     "note": "Toxicity labelling."
    },
    {
     "name": "language_identifier",
     "url": "https://github.com/latam-gpt/language_identifier",
     "note": "Language identifier."
    },
    {
     "name": "training-estimations",
     "url": "https://github.com/latam-gpt/training-estimations",
     "note": "Training estimations."
    }
   ],
   "hf_highlights": [
    {
     "name": "Llama-3.1-70B-LatamGPT-SFT-1.0",
     "url": "https://huggingface.co/latam-gpt/Llama-3.1-70B-LatamGPT-SFT-1.0",
     "kind": "model",
     "note": "SFT 70B."
    },
    {
     "name": "Wayra-Perplexity-Estimator-55M",
     "url": "https://huggingface.co/latam-gpt/Wayra-Perplexity-Estimator-55M",
     "kind": "model",
     "note": "Perplexity estimator."
    },
    {
     "name": "LatamGPT-Corpus-1.0",
     "url": "https://huggingface.co/datasets/latam-gpt/LatamGPT-Corpus-1.0",
     "kind": "dataset",
     "note": "Corpus."
    },
    {
     "name": "Trueque-Benchmark-beta-0.1",
     "url": "https://huggingface.co/datasets/latam-gpt/Trueque-Benchmark-beta-0.1",
     "kind": "dataset",
     "note": "Benchmark."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Llama-3.1-70B-LatamGPT-SFT-1.0",
     "url": "https://huggingface.co/latam-gpt/Llama-3.1-70B-LatamGPT-SFT-1.0",
     "note": "Instruction-tuned model"
    },
    {
     "name": "LatamGPT-Corpus-1.0",
     "url": "https://huggingface.co/datasets/latam-gpt/LatamGPT-Corpus-1.0",
     "note": "Regional corpus"
    },
    {
     "name": "Wayra Perplexity Estimator",
     "url": "https://huggingface.co/latam-gpt/Wayra-Perplexity-Estimator-55M",
     "note": "Data quality model"
    },
    {
     "name": "Trueque benchmark",
     "url": "https://huggingface.co/datasets/latam-gpt/Trueque-Benchmark-beta-0.1",
     "note": "Benchmark"
    }
   ],
   "openness": "Weights, corpus and tools are public on Hugging Face and GitHub; the SFT model carries a Llama 3.1 licence and the corpus a custom licence.",
   "key_people": [
    {
     "name": "Álvaro Soto",
     "role": "Director of CENIA",
     "url": "https://www.intellinews.com/chile-launches-latam-gpt-in-push-for-regional-ai-sovereignty-425563/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://latamgpt.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://latamgpt.org",
    "https://cenia.cl",
    "https://huggingface.co/latam-gpt",
    "https://github.com/latam-gpt"
   ],
   "confidence": "high",
   "region": "Latin America",
   "id": "latam-gpt",
   "logo": true
  },
  {
   "orig_id": "lelapa-ai",
   "name": "Lelapa AI",
   "alias": null,
   "full_name": "Lelapa AI",
   "type": "startup",
   "founded": 2022,
   "country": "South Africa",
   "city": "Johannesburg",
   "lat": -26.2041,
   "lon": 28.0473,
   "website": "https://lelapa.ai",
   "huggingface": "https://huggingface.co/lelapa",
   "github": "https://github.com/lelapa-ai",
   "x": "https://twitter.com/LelapaAI",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/lelapa-ai/",
   "contact_email": "support@lelapa.ai",
   "summary": "Lelapa AI is a South African startup building language AI for African languages, including a commercial speech and translation product called Vulavula. It also releases research models and datasets openly, notably InkubaLM.",
   "tagline": "South African language AI company building resource-efficient models for African languages, maker of Vulavula and the open InkubaLM.",
   "about": [
    "Lelapa AI was founded in Johannesburg in 2022 by Pelonomi Moiloa, Jade Abbott and academic collaborators, and the name means home in Sotho and Tswana. Moiloa is chief executive and Abbott is chief technology officer, and the team page also lists Vukosi Marivate as chief information officer, Benjamin Rosman as chief scientist and George Konidaris as an advisor. Brown University and Wired have described Moiloa, Abbott, Marivate, Konidaris, Rosman and Pravesh Ranchod as the founding group.",
    "The company's commercial product line is Vulavula, with Vulavula Transcribe for multilingual contact centres and Vulavula Translate for translation between African languages. It targets telecoms and financial services.",
    "Lelapa also releases research models and datasets openly, notably InkubaLM, and states that it builds resource-efficient language AI designed to scale."
   ],
   "mission": "Lelapa states that it builds resource-efficient language AI for global scale, starting with African languages that are poorly served by existing models.",
   "notice": null,
   "known_for": [
    "Vulavula speech and translation products",
    "InkubaLM small African-language model",
    "Founded by Masakhane and academic researchers"
   ],
   "focus": [
    "language models",
    "speech",
    "low-resource languages",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Lelapa AI founded",
     "note": "Founded in Johannesburg to build language AI for African languages.",
     "year": "2022"
    },
    {
     "title": "Vulavula",
     "note": "Launched speech and translation services for African languages.",
     "year": "2022"
    },
    {
     "title": "InkubaLM",
     "note": "Released a small open multilingual model for African languages.",
     "year": "2024"
    },
    {
     "title": "Vulavula Transcribe and Translate",
     "note": "Current product line aimed at contact centres and translation between local languages.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "zindi-inkuba-notebook",
     "url": "https://github.com/lelapa-ai/zindi-inkuba-notebook",
     "note": "Notebook for the Inkuba challenge."
    },
    {
     "name": "vulavula-examples",
     "url": "https://github.com/lelapa-ai/vulavula-examples",
     "note": "Vulavula examples."
    },
    {
     "name": "lelapa-demo-gnerds",
     "url": "https://github.com/lelapa-ai/lelapa-demo-gnerds",
     "note": "Translation demo."
    }
   ],
   "hf_highlights": [
    {
     "name": "InkubaLM-0.4B",
     "url": "https://huggingface.co/lelapa/InkubaLM-0.4B",
     "kind": "model",
     "note": "InkubaLM."
    },
    {
     "name": "Inkuba-instruct",
     "url": "https://huggingface.co/datasets/lelapa/Inkuba-instruct",
     "kind": "dataset",
     "note": "Instruction data."
    },
    {
     "name": "Inkuba-Mono",
     "url": "https://huggingface.co/datasets/lelapa/Inkuba-Mono",
     "kind": "dataset",
     "note": "Monolingual data."
    },
    {
     "name": "Vukuzenzele_isiXhosa_Speech_Dataset_ViXSD",
     "url": "https://huggingface.co/datasets/lelapa/Vukuzenzele_isiXhosa_Speech_Dataset_ViXSD",
     "kind": "dataset",
     "note": "isiXhosa speech."
    }
   ],
   "papers": [
    {
     "title": "InkubaLM: A small language model for low-resource African languages",
     "url": "https://arxiv.org/abs/2408.17024",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "InkubaLM-0.4B",
     "url": "https://huggingface.co/lelapa/InkubaLM-0.4B",
     "note": "Small language model for African languages"
    },
    {
     "name": "Inkuba-instruct",
     "url": "https://huggingface.co/datasets/lelapa/Inkuba-instruct",
     "note": "Instruction dataset"
    },
    {
     "name": "Inkuba-Mono",
     "url": "https://huggingface.co/datasets/lelapa/Inkuba-Mono",
     "note": "Monolingual dataset"
    },
    {
     "name": "ViXSD isiXhosa speech",
     "url": "https://huggingface.co/datasets/lelapa/Vukuzenzele_isiXhosa_Speech_Dataset_ViXSD",
     "note": "isiXhosa speech dataset"
    }
   ],
   "openness": "Research models and datasets are on Hugging Face (InkubaLM-0.4B is listed under CC BY-NC 4.0); the commercial Vulavula product is closed.",
   "key_people": [
    {
     "name": "Pelonomi Moiloa",
     "role": "Chief Executive Officer",
     "url": "https://lelapa.ai/team",
     "note": null
    },
    {
     "name": "Jade Abbott",
     "role": "Chief Technology Officer",
     "url": "https://lelapa.ai/team",
     "note": null
    },
    {
     "name": "Vukosi Marivate",
     "role": "Chief Information Officer",
     "url": "https://lelapa.ai/team",
     "note": null
    },
    {
     "name": "Benjamin Rosman",
     "role": "Chief Scientist",
     "url": "https://lelapa.ai/team",
     "note": null
    },
    {
     "name": "George Konidaris",
     "role": "Advisor",
     "url": "https://posts.cs.brown.edu/2023/03/31/george-konidaris-cofounds-lelapa-develop-homegrown-african-ai/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://lelapa.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://lelapa.ai",
    "https://huggingface.co/lelapa",
    "https://github.com/Lelapa-AI",
    "https://arxiv.org/abs/2408.17024"
   ],
   "confidence": "high",
   "region": "Africa",
   "id": "lelapa-ai",
   "logo": true
  },
  {
   "orig_id": "letta",
   "name": "Letta",
   "alias": null,
   "full_name": "Letta",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.letta.com",
   "huggingface": "https://huggingface.co/letta-ai",
   "github": "https://github.com/letta-ai",
   "x": null,
   "discord": "https://discord.gg/letta",
   "linkedin": null,
   "contact_email": null,
   "summary": "San Francisco AI research lab from the creators of MemGPT, building open-source stateful agents.",
   "tagline": "San Francisco AI lab from the creators of MemGPT, building open-source stateful agents that remember and learn over time.",
   "about": [
    "Letta grew out of MemGPT, a UC Berkeley Sky Computing Lab project on virtual context management for language model agents. Founders Charles Packer, chief executive, and Sarah Wooders, chief technology officer, met during their PhD research there under Ion Stoica and Joseph Gonzalez, who advise the company.",
    "In September 2024 the company emerged from stealth with a 10 million dollar seed round led by Felicis, with Sunflower Capital and Essence VC participating. It released an Agent Development Environment and API platform alongside its open-source framework.",
    "Letta describes itself as a research lab building machines that learn, aimed at agents that remember everything and improve continuously."
   ],
   "mission": "Letta states that it is building machines that learn: experiential agents that remember, learn continuously and improve themselves over time.",
   "notice": null,
   "known_for": [
    "MemGPT agent memory research",
    "Open-source Letta agent framework",
    "Stateful long-running agents"
   ],
   "focus": [
    "agents",
    "memory"
   ],
   "contributions": [
    {
     "title": "MemGPT paper",
     "note": "Published MemGPT, treating a language model's context as virtual memory managed by the agent.",
     "year": "2023"
    },
    {
     "title": "Letta launch and seed round",
     "note": "Emerged from stealth with a 10 million dollar seed round led by Felicis.",
     "year": "2024"
    },
    {
     "title": "Agent Development Environment",
     "note": "Released an ADE and API platform for building stateful agents.",
     "year": "2024"
    },
    {
     "title": "Sleep-time compute",
     "note": "Published research on letting agents use idle time to pre-process memory.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "letta",
     "url": "https://github.com/letta-ai/letta",
     "note": "Platform for stateful agents."
    },
    {
     "name": "letta-code",
     "url": "https://github.com/letta-ai/letta-code",
     "note": "Self-improving coding agents."
    },
    {
     "name": "claude-subconscious",
     "url": "https://github.com/letta-ai/claude-subconscious",
     "note": "Memory layer for Claude Code."
    },
    {
     "name": "agent-file",
     "url": "https://github.com/letta-ai/agent-file",
     "note": "Agent File format."
    }
   ],
   "hf_highlights": [
    {
     "name": "frontier-memory-corpus",
     "url": "https://huggingface.co/datasets/letta-ai/frontier-memory-corpus",
     "kind": "dataset",
     "note": "Memory corpus."
    },
    {
     "name": "SWE-Features",
     "url": "https://huggingface.co/datasets/letta-ai/SWE-Features",
     "kind": "dataset",
     "note": "SWE features."
    },
    {
     "name": "stateful-gsm-symbolic",
     "url": "https://huggingface.co/datasets/letta-ai/stateful-gsm-symbolic",
     "kind": "dataset",
     "note": "Stateful GSM-Symbolic."
    }
   ],
   "papers": [
    {
     "title": "MemGPT: Towards LLMs as Operating Systems",
     "url": "https://arxiv.org/abs/2310.08560",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "letta",
     "url": "https://github.com/letta-ai/letta",
     "note": "Platform"
    },
    {
     "name": "agent-file",
     "url": "https://github.com/letta-ai/agent-file",
     "note": "Agent file format"
    }
   ],
   "openness": "Open-source code.",
   "key_people": [
    {
     "name": "Charles Packer",
     "role": "Co-founder and CEO",
     "url": "https://techcrunch.com/2024/09/23/letta-one-of-uc-berkeleys-most-anticipated-ai-startups-has-just-come-out-of-stealth",
     "note": null
    },
    {
     "name": "Sarah Wooders",
     "role": "Co-founder and CTO",
     "url": "https://techcrunch.com/2024/09/23/letta-one-of-uc-berkeleys-most-anticipated-ai-startups-has-just-come-out-of-stealth",
     "note": null
    },
    {
     "name": "Ion Stoica",
     "role": "Advisor",
     "url": "https://www.letta.com/about-us",
     "note": null
    },
    {
     "name": "Joseph Gonzalez",
     "role": "Advisor",
     "url": "https://www.letta.com/about-us",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.letta.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.letta.com/about-us",
    "https://github.com/letta-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "letta",
   "mlat": 37.8942,
   "mlon": -122.43581,
   "logo": true
  },
  {
   "orig_id": "lightblue",
   "name": "Lightblue",
   "alias": null,
   "full_name": "Lightblue",
   "type": "startup",
   "founded": null,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://www.lightblue-tech.com/",
   "huggingface": "https://huggingface.co/lightblue",
   "github": "https://github.com/lightblue-tech",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Tokyo firm that releases open models such as rerankers on Hugging Face.",
   "tagline": "Tokyo AI company that publishes open Japanese and multilingual language models on Hugging Face alongside its enterprise AI platform.",
   "about": [
    "Lightblue is a Tokyo company based in the Otemachi district. Its commercial offering is an AI platform for daily work that lets organizations build assistants on their own knowledge, together with professional services for AI implementation and video analysis.",
    "On Hugging Face, Lightblue KK publishes open models. These include the Suzume series, finetunes of Llama 3 8B for Japanese and for multiple languages, developed by Peter Devine, and named after the Japanese tree sparrow.",
    "The company's own pages do not describe an open research agenda, so its open-source work is mainly visible through its model releases."
   ],
   "mission": "Lightblue's stated business goal is to help organizations settle AI into daily operations; its model releases give Japanese developers open options.",
   "notice": null,
   "known_for": [
    "Suzume Japanese and multilingual models",
    "Open model releases on Hugging Face",
    "Enterprise AI assistants in Japan"
   ],
   "focus": [
    "language models",
    "reranking",
    "Japanese"
   ],
   "contributions": [
    {
     "title": "Suzume Llama 3 8B multilingual",
     "note": "Published an open multilingual finetune of Llama 3 8B.",
     "year": "2024"
    },
    {
     "title": "Suzume Llama 3 8B Japanese",
     "note": "Published a Japanese finetune that Lightblue reports as best in the 7B and 8B class on Japanese benchmarks.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "lb-reranker",
     "url": "https://github.com/lightblue-tech/lb-reranker",
     "note": "Reranker."
    },
    {
     "name": "M-IFEval",
     "url": "https://github.com/lightblue-tech/M-IFEval",
     "note": "Multilingual instruction-following evaluation."
    },
    {
     "name": "shitsu",
     "url": "https://github.com/lightblue-tech/shitsu",
     "note": "Fast multilingual text quality classifier."
    },
    {
     "name": "aloftrag",
     "url": "https://github.com/lightblue-tech/aloftrag",
     "note": "ALoFTRAG code."
    }
   ],
   "hf_highlights": [
    {
     "name": "suzume-llama-3-8B-multilingual-orpo-borda-top25",
     "url": "https://huggingface.co/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top25",
     "kind": "model",
     "note": "Suzume variant."
    },
    {
     "name": "suzume-llama-3-8B-multilingual",
     "url": "https://huggingface.co/lightblue/suzume-llama-3-8B-multilingual",
     "kind": "model",
     "note": "Suzume base."
    },
    {
     "name": "rag_multilingual_training_negatives",
     "url": "https://huggingface.co/datasets/lightblue/rag_multilingual_training_negatives",
     "kind": "dataset",
     "note": "RAG data."
    },
    {
     "name": "reasoning-multilingual-R1-Llama-70B-train",
     "url": "https://huggingface.co/datasets/lightblue/reasoning-multilingual-R1-Llama-70B-train",
     "kind": "dataset",
     "note": "Reasoning data."
    }
   ],
   "papers": [
    {
     "title": "ALoFTRAG: Automatic Local Fine Tuning for Retrieval Augmented Generation",
     "url": "https://arxiv.org/abs/2501.11929",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "lb-reranker-0.5B-v1.0",
     "url": "https://huggingface.co/lightblue/lb-reranker-0.5B-v1.0",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [
    {
     "name": "Peter Devine",
     "role": "Developer of the Suzume models",
     "url": "https://huggingface.co/lightblue/suzume-llama-3-8B-japanese",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website.",
   "sources": [
    "https://huggingface.co/lightblue",
    "https://www.lightblue-tech.com/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "lightblue",
   "mlat": 35.64938,
   "mlon": 139.75529,
   "logo": true
  },
  {
   "orig_id": "lightly",
   "name": "Lightly",
   "alias": null,
   "full_name": "Lightly",
   "type": "startup",
   "founded": null,
   "country": "Switzerland",
   "city": "Zurich",
   "lat": 47.3769,
   "lon": 8.5417,
   "website": "https://www.lightly.ai",
   "huggingface": "https://huggingface.co/lightly-ai",
   "github": "https://github.com/lightly-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Zurich startup open-sourcing self-supervised learning and vision training libraries.",
   "tagline": "Lightly is a Zurich company, born from ETH research, that open-sources self-supervised learning and vision training libraries.",
   "about": [
    "Lightly was founded by Matthias Heller and Igor Susmelj as a data-centric AI company based on ETH Zurich research. It raised a 3 million dollar seed round in May 2022 and reports 10 million dollars raised in total.",
    "Its open-source lightly library for self-supervised learning dates from October 2020. It later added the labelformat converter, LightlyTrain for vision model pretraining and distillation, and LightlyStudio, alongside a commercial platform."
   ],
   "mission": "To help teams build better vision models by improving the data that models are trained on.",
   "notice": "Lightly is a commercial startup. Some of its research code and models are open source.",
   "known_for": [
    "Lightly self-supervised learning library",
    "LightlyTrain vision pretraining",
    "ETH Zurich spin-off"
   ],
   "focus": [
    "computer vision",
    "self-supervised learning"
   ],
   "contributions": [
    {
     "title": "lightly library",
     "note": "The open-source self-supervised learning library was published on GitHub.",
     "year": "2020"
    },
    {
     "title": "Seed round",
     "note": "Lightly secured 3 million dollars led by Wingman Ventures.",
     "year": "2022"
    },
    {
     "title": "labelformat",
     "note": "A converter between computer vision label formats was released.",
     "year": "2023"
    },
    {
     "title": "LightlyTrain",
     "note": "A training toolkit for YOLO, ViTs and RT-DETR was released.",
     "year": "2025"
    },
    {
     "title": "LightlyStudio",
     "note": "A unified data platform for multimodal ML was published.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "lightly",
     "url": "https://github.com/lightly-ai/lightly",
     "note": "A python library for self-supervised learning and latent world models on images. (3,832 stars)"
    },
    {
     "name": "lightly-train",
     "url": "https://github.com/lightly-ai/lightly-train",
     "note": "All-in-one training for vision models (YOLO, ViTs, RT-DETR, DINOv3): pretraining, fine-tuning, distillation. (1,741 stars)"
    },
    {
     "name": "lightly-studio",
     "url": "https://github.com/lightly-ai/lightly-studio",
     "note": "LightlyStudio - The Unified Data Platform for Multimodal ML (911 stars)"
    },
    {
     "name": "labelformat",
     "url": "https://github.com/lightly-ai/labelformat",
     "note": "A tool for converting computer vision label formats. (91 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "lightly-ai/simclrv1-imagenet1k-resnet50-4x",
     "url": "https://huggingface.co/lightly-ai/simclrv1-imagenet1k-resnet50-4x",
     "kind": "model",
     "note": "Model with 1 likes on the Hub."
    },
    {
     "name": "lightly-ai/simclrv1-imagenet1k-resnet50-1x",
     "url": "https://huggingface.co/lightly-ai/simclrv1-imagenet1k-resnet50-1x",
     "kind": "model",
     "note": "Model with 0 likes on the Hub."
    },
    {
     "name": "lightly-ai/epic-kitchens-100-clips",
     "url": "https://huggingface.co/datasets/lightly-ai/epic-kitchens-100-clips",
     "kind": "dataset",
     "note": "Dataset with 4 likes on the Hub."
    },
    {
     "name": "lightly-ai/SAMv2-Mask-Generator",
     "url": "https://huggingface.co/spaces/lightly-ai/SAMv2-Mask-Generator",
     "kind": "space",
     "note": "Space with 18 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "lightly",
     "url": "https://github.com/lightly-ai/lightly",
     "note": "Self-supervised learning library (3,800+ stars)"
    },
    {
     "name": "LightlyTrain",
     "url": "https://github.com/lightly-ai/lightly-train",
     "note": "Pretraining and fine-tuning for vision models"
    },
    {
     "name": "LightlyStudio",
     "url": "https://github.com/lightly-ai/lightly-studio",
     "note": "Data platform for multimodal ML"
    },
    {
     "name": "labelformat",
     "url": "https://github.com/lightly-ai/labelformat",
     "note": "Label converter"
    }
   ],
   "openness": "Open-source libraries with a commercial platform alongside.",
   "key_people": [
    {
     "name": "Matthias Heller",
     "role": "Co-founder and CEO",
     "url": "https://www.lightly.ai/about",
     "note": null
    },
    {
     "name": "Igor Susmelj",
     "role": "Co-founder and CTO",
     "url": "https://huggingface.co/igor-lightly",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://www.lightly.ai",
    "https://huggingface.co/lightly-ai",
    "https://github.com/lightly-ai"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "lightly",
   "mlat": 47.3336,
   "mlon": 8.5417,
   "logo": true
  },
  {
   "orig_id": "lightning-ai",
   "name": "Lightning AI",
   "alias": null,
   "full_name": "Lightning AI",
   "type": "startup",
   "founded": 2019,
   "country": "United States",
   "city": "New York",
   "lat": 40.7128,
   "lon": -74.006,
   "website": "https://lightning.ai",
   "huggingface": null,
   "github": "https://github.com/Lightning-AI",
   "x": null,
   "discord": "https://discord.gg/MWAEvnC5fU",
   "linkedin": null,
   "contact_email": null,
   "summary": "Company behind PyTorch Lightning, LitGPT and LitServe, open-source frameworks for training and serving models.",
   "tagline": "Company behind PyTorch Lightning, LitGPT and LitServe, open-source frameworks for training and serving models.",
   "about": [
    "Lightning AI was founded by William Falcon, who began developing PyTorch Lightning in 2015 at Columbia University and open-sourced it in 2019 while completing a PhD at NYU and working at Facebook AI Research. His advisors were Kyunghyun Cho and Yann LeCun.",
    "The company's open-source tools include PyTorch Lightning, Fabric, LitGPT, LitServe, TorchMetrics and the Thunder compiler. In 2023 it launched Lightning AI Studio, a browser-based cloud platform for coding, training and deploying models.",
    "In 2026 Lightning AI and the GPU cloud provider Voltage Park combined into one company, offering a platform that spans GPUs to developer tools."
   ],
   "mission": "Lightning AI aims to give developers one platform from hardware to tooling for building AI, while maintaining its widely used open-source frameworks.",
   "notice": null,
   "known_for": [
    "PyTorch Lightning",
    "LitGPT and LitServe",
    "Lightning AI Studio"
   ],
   "focus": [
    "training infrastructure",
    "libraries",
    "language models"
   ],
   "contributions": [
    {
     "title": "PyTorch Lightning open-sourced",
     "note": "Falcon released the training framework that became widely used in research.",
     "year": "2019"
    },
    {
     "title": "Lightning AI Studio",
     "note": "Launched a browser-based cloud platform for coding, training and deploying models.",
     "year": "2023"
    },
    {
     "title": "LitGPT",
     "note": "Published an open implementation library for pretraining and fine-tuning LLMs.",
     "year": "2023"
    },
    {
     "title": "LitServe",
     "note": "Released an open framework for serving models.",
     "year": "2024"
    },
    {
     "title": "Merger with Voltage Park",
     "note": "Combined with Voltage Park into a single company covering GPUs through developer tools.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "pytorch-lightning",
     "url": "https://github.com/Lightning-AI/pytorch-lightning",
     "note": "Training framework."
    },
    {
     "name": "litgpt",
     "url": "https://github.com/Lightning-AI/litgpt",
     "note": "LLM recipes."
    },
    {
     "name": "lit-llama",
     "url": "https://github.com/Lightning-AI/lit-llama",
     "note": "LLaMA implementation."
    },
    {
     "name": "LitServe",
     "url": "https://github.com/Lightning-AI/LitServe",
     "note": "Serving framework."
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "pytorch-lightning",
     "url": "https://github.com/Lightning-AI/pytorch-lightning",
     "note": "Training framework"
    },
    {
     "name": "litgpt",
     "url": "https://github.com/Lightning-AI/litgpt",
     "note": "LLM recipes"
    },
    {
     "name": "LitServe",
     "url": "https://github.com/Lightning-AI/LitServe",
     "note": "Serving"
    }
   ],
   "openness": "Open-source Apache-2.0 libraries; commercial cloud alongside.",
   "key_people": [
    {
     "name": "William Falcon",
     "role": "Founder and CEO",
     "url": "https://lightning.ai/about",
     "note": "Creator of PyTorch Lightning."
    },
    {
     "name": "Ozan Kaya",
     "role": "President; former CEO of Voltage Park",
     "url": "https://www.businesswire.com/news/home/20260121371691/en/Lightning-AI-and-Voltage-Park-Complete-Merger-to-Create-the-First-Cloud-Built-for-AI",
     "note": "Quoted in the January 2026 merger announcement."
    },
    {
     "name": "Saurabh Giri",
     "role": "Chief Product and Technology Officer",
     "url": "https://www.businesswire.com/news/home/20260121371691/en/Lightning-AI-and-Voltage-Park-Complete-Merger-to-Create-the-First-Cloud-Built-for-AI",
     "note": "Named in the merger announcement."
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://lightning.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://lightning.ai",
    "https://github.com/Lightning-AI"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "lightning-ai",
   "mlat": 40.69553,
   "mlon": -73.93586,
   "logo": true
  },
  {
   "orig_id": "lighton",
   "name": "LightOn",
   "alias": null,
   "full_name": "LightOn",
   "type": "startup",
   "founded": 2016,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.lighton.ai",
   "huggingface": "https://huggingface.co/lightonai",
   "github": "https://github.com/lightonai",
   "x": "https://x.com/LightOnIO",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/lighton",
   "contact_email": "contact@lighton.ai",
   "summary": "Paris-based AI company working on enterprise search and document understanding. It open-sources retrieval models, OCR models and multi-vector search tools.",
   "tagline": "Paris-based, publicly listed AI company that open-sources retrieval models, ModernBERT variants and multi-vector search tools.",
   "about": [
    "LightOn was founded in Paris in 2016 by Igor Carron, Laurent Daudet, Florent Krzakala and Sylvain Gigan. Carron is chief executive and Daudet is deputy managing director. It launched an optical processing unit in 2018, and an OPU was integrated into the Jean Zay supercomputer in 2021.",
    "The company began training language models in 2020 and says it has trained more than twelve since. It has open-sourced PAGnol for French, RITA for proteins, Mambaoutai, and models built with AnswerAI such as ModernBERT, plus retrieval models like GTE-ModernColBERT and tools such as PyLate and FastPlaid.",
    "In November 2024 LightOn listed on Euronext Growth Paris and describes itself as Europe's first listed generative AI startup. Its commercial product is Paradigm, an on-premise enterprise platform."
   ],
   "mission": "LightOn builds enterprise search and document understanding on private infrastructure while publishing retrieval models and tools as open source.",
   "notice": null,
   "known_for": [
    "ModernBERT with AnswerAI",
    "PyLate multi-vector retrieval",
    "PAGnol French language models"
   ],
   "focus": [
    "retrieval",
    "OCR",
    "embeddings",
    "language models"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in Paris by four scientists.",
     "year": "2016"
    },
    {
     "title": "Optical Processing Unit",
     "note": "Launched an optical hardware accelerator.",
     "year": "2018"
    },
    {
     "title": "First open LLMs",
     "note": "Began training and releasing language models, including PAGnol.",
     "year": "2020"
    },
    {
     "title": "Jean Zay integration",
     "note": "An OPU was integrated into a French national supercomputer.",
     "year": "2021"
    },
    {
     "title": "IPO and ModernBERT",
     "note": "Listed on Euronext Growth Paris and released ModernBERT.",
     "year": "2024"
    },
    {
     "title": "Ettin Suite",
     "note": "Released a family of encoder and decoder models.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "pylate",
     "url": "https://github.com/lightonai/pylate",
     "note": "Late Interaction Models Training & Retrieval (910 stars)"
    },
    {
     "name": "next-plaid",
     "url": "https://github.com/lightonai/next-plaid",
     "note": "NextPlaid, ColGREP: Multi-vector search, from database to coding agents. (619 stars)"
    },
    {
     "name": "fast-plaid",
     "url": "https://github.com/lightonai/fast-plaid",
     "note": "High-Performance Engine for Multi-Vector Search (286 stars)"
    },
    {
     "name": "RITA",
     "url": "https://github.com/lightonai/RITA",
     "note": "RITA is a family of autoregressive protein models, developed by LightOn in collaboration with the OATML group  (100 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "lightonai/LightOnOCR-2-1B",
     "url": "https://huggingface.co/lightonai/LightOnOCR-2-1B",
     "kind": "model",
     "note": "Model: 803,844 downloads, 852 likes."
    },
    {
     "name": "lightonai/GTE-ModernColBERT-v1",
     "url": "https://huggingface.co/lightonai/GTE-ModernColBERT-v1",
     "kind": "model",
     "note": "Model: 16,574 downloads, 180 likes."
    },
    {
     "name": "lightonai/multilingual-embeddings-pre-training-curated",
     "url": "https://huggingface.co/datasets/lightonai/multilingual-embeddings-pre-training-curated",
     "kind": "dataset",
     "note": "Dataset: 16,558 downloads, 5 likes."
    },
    {
     "name": "lightonai/LightOnOCR-2-1B-Demo",
     "url": "https://huggingface.co/spaces/lightonai/LightOnOCR-2-1B-Demo",
     "kind": "space",
     "note": "Space: 129 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "PyLate: Flexible Training and Retrieval for Late Interaction Models",
     "url": "https://arxiv.org/abs/2508.03555",
     "year": "2025"
    },
    {
     "title": "RITA: a Study on Scaling Up Generative Protein Sequence Models",
     "url": "https://arxiv.org/abs/2205.05789",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "LightOnOCR-2-1B",
     "url": "https://huggingface.co/lightonai/LightOnOCR-2-1B",
     "note": "Compact OCR vision-language model"
    },
    {
     "name": "PyLate",
     "url": "https://github.com/lightonai/pylate",
     "note": "Library for late-interaction retrieval models"
    },
    {
     "name": "Reason-ModernColBERT",
     "url": "https://huggingface.co/lightonai/Reason-ModernColBERT",
     "note": "Reasoning-oriented ColBERT retriever"
    },
    {
     "name": "FastPlaid",
     "url": "https://github.com/lightonai/fast-plaid",
     "note": "Multi-vector search engine"
    }
   ],
   "openness": "Publishes models, datasets and libraries on Hugging Face and GitHub, mostly under MIT or Apache-2.0 for code; check model cards for model licences.",
   "key_people": [
    {
     "name": "Igor Carron",
     "role": "Co-founder and CEO",
     "url": "https://en.wikipedia.org/wiki/LightOn",
     "note": null
    },
    {
     "name": "Laurent Daudet",
     "role": "Co-founder and deputy managing director",
     "url": "https://en.wikipedia.org/wiki/LightOn",
     "note": null
    },
    {
     "name": "Florent Krzakala",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/LightOn",
     "note": null
    },
    {
     "name": "Sylvain Gigan",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/LightOn",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact address printed on the website.",
   "sources": [
    "https://www.lighton.ai",
    "https://huggingface.co/lightonai",
    "https://github.com/lightonai",
    "https://arxiv.org/abs/2508.03555"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "lighton",
   "mlat": 48.95792,
   "mlon": 2.38099,
   "logo": true
  },
  {
   "orig_id": "linagora",
   "name": "LINAGORA",
   "alias": null,
   "full_name": "LINAGORA",
   "type": "startup",
   "founded": 2000,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://linagora.com",
   "huggingface": "https://huggingface.co/linagora",
   "github": "https://github.com/linagora",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "French open-source software company behind the LinTO speech stack and partner in the Lucie French-focused open LLM project.",
   "tagline": "LINAGORA is a French open-source software company behind the LinTO voice assistant and the Lucie French open LLM.",
   "about": [
    "LINAGORA was founded in 2000 by Alexandre Zapolsky and Michel-Marie Maudet, and its headquarters moved to Issy-les-Moulineaux in 2024. Its products include Twake Workplace, OpenPaaS and the LinTO voice assistant, and its GitHub organization holds more than 500 repositories.",
    "In 2023 it created the OpenLLM France community, and in January 2025 it released Lucie, an open-source sovereign large language model. The release drew criticism after tests of an unfinished version and was temporarily suspended."
   ],
   "mission": "To invent and develop ethical, open-source technology and to support open sovereign AI in France.",
   "notice": null,
   "known_for": [
    "LinTO open-source voice assistant",
    "Lucie French open LLM",
    "Twake Workplace and OpenPaaS"
   ],
   "focus": [
    "speech recognition",
    "LLMs",
    "open-source software"
   ],
   "contributions": [
    {
     "title": "French National Assembly adopts its software",
     "note": "The National Assembly chose Linagora software to replace Windows.",
     "year": "2007"
    },
    {
     "title": "LinTO launched",
     "note": "The open-source voice assistant was launched.",
     "year": "2017"
    },
    {
     "title": "OpenLLM France",
     "note": "Linagora created the community around open French language models.",
     "year": "2023"
    },
    {
     "title": "CNRS framework agreement",
     "note": "Linagora and CNRS signed a research collaboration agreement.",
     "year": "2024"
    },
    {
     "title": "Lucie released",
     "note": "The open-source French LLM was released in January 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "hublin",
     "url": "https://github.com/linagora/hublin",
     "note": "DEPRECATED - An easy and free video conference service based on WebRTC (1,603 stars)"
    },
    {
     "name": "linshare",
     "url": "https://github.com/linagora/linshare",
     "note": "LinShare - Secure File Sharing (582 stars)"
    },
    {
     "name": "openpaas-esn",
     "url": "https://github.com/linagora/openpaas-esn",
     "note": "Open PaaS Enterprise Social Network (485 stars)"
    },
    {
     "name": "jmap-client",
     "url": "https://github.com/linagora/jmap-client",
     "note": "This repository provides a javascript lib to make requests against a JMAP server. (72 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "linagora/linto-asr-ar-tn-0.1",
     "url": "https://huggingface.co/linagora/linto-asr-ar-tn-0.1",
     "kind": "model",
     "note": "Model with 16 likes on the Hub."
    },
    {
     "name": "linagora/linto_stt_fr_fastconformer",
     "url": "https://huggingface.co/linagora/linto_stt_fr_fastconformer",
     "kind": "model",
     "note": "Model with 8 likes on the Hub."
    },
    {
     "name": "linagora/linto-dataset-audio-ar-tn",
     "url": "https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn",
     "kind": "dataset",
     "note": "Dataset with 22 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "The Lucie-7B LLM and the Lucie Training Dataset: Open resources for multilingual language generation",
     "url": "https://arxiv.org/abs/2503.12294",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "LinTO ASR models",
     "url": "https://huggingface.co/linagora/linto_stt_fr_fastconformer",
     "note": "French speech recognition model"
    },
    {
     "name": "Labess-7b-chat",
     "url": "https://huggingface.co/linagora/Labess-7b-chat",
     "note": "Tunisian Arabic chat model"
    },
    {
     "name": "SUMM-RE",
     "url": "https://huggingface.co/datasets/linagora/SUMM-RE",
     "note": "French meeting speech dataset"
    },
    {
     "name": "LinShare",
     "url": "https://github.com/linagora/linshare",
     "note": "Open-source secure file sharing"
    }
   ],
   "openness": "Long-standing open-source publisher (GitHub and Hugging Face), including speech models and datasets under open licences.",
   "key_people": [
    {
     "name": "Alexandre Zapolsky",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Linagora",
     "note": null
    },
    {
     "name": "Michel-Marie Maudet",
     "role": "Co-founder and CEO",
     "url": "https://github.com/mmaudet",
     "note": null
    },
    {
     "name": "Julie Hunter",
     "role": "Researcher",
     "url": "https://huggingface.co/juliehunter",
     "note": null
    },
    {
     "name": "Olivier Gouvert",
     "role": "Researcher",
     "url": "https://huggingface.co/oligou",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://linagora.com",
    "https://huggingface.co/linagora",
    "https://github.com/linagora",
    "https://arxiv.org/abs/2503.12294"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "linagora",
   "mlat": 48.95792,
   "mlon": 2.32341,
   "logo": true
  },
  {
   "orig_id": "liquid-ai",
   "name": "Liquid AI",
   "alias": null,
   "full_name": "Liquid AI",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "Cambridge",
   "lat": 42.3736,
   "lon": -71.1097,
   "website": "https://www.liquid.ai",
   "huggingface": "https://huggingface.co/LiquidAI",
   "github": "https://github.com/Liquid4All",
   "x": "https://x.com/liquidai",
   "discord": "https://discord.com/invite/liquid-ai",
   "linkedin": null,
   "contact_email": null,
   "summary": "MIT spin-out building efficient on-device foundation models (LFM) with open-weight releases.",
   "tagline": "MIT CSAIL spin-out building efficient on-device foundation models, the LFM family, with open weights on Hugging Face.",
   "about": [
    "Liquid AI was spun out of MIT's Computer Science and Artificial Intelligence Laboratory by Ramin Hasani, chief executive, Mathias Lechner, chief technology officer, Alexander Amini, chief scientist, and Daniela Rus. Reports date its start to 2023.",
    "It builds Liquid Foundation Models, designed for efficient inference on phones, laptops and other edge devices. It released LFM2 in July 2025 as small open-weight models, followed by a technical report in December 2025 and the LFM2.5 and LFM2-24B-A2B models.",
    "Weights are published on Hugging Face, and the company also sells on-premise and edge deployments."
   ],
   "mission": "Liquid AI aims to build efficient general-purpose models that run on-device and on-premise, and to release open-weight versions for developers.",
   "notice": null,
   "known_for": [
    "LFM2 on-device foundation models",
    "Liquid neural network research",
    "Efficient small open-weight models"
   ],
   "focus": [
    "language models",
    "edge AI",
    "audio"
   ],
   "contributions": [
    {
     "title": "Company started",
     "note": "Founded by MIT CSAIL researchers.",
     "year": "2023"
    },
    {
     "title": "First LFM series",
     "note": "Announced the first Liquid Foundation Models.",
     "year": "2024"
    },
    {
     "title": "LFM2",
     "note": "Released open-weight small models of 0.35B, 0.7B and 1.2B parameters.",
     "year": "2025"
    },
    {
     "title": "LFM2 technical report",
     "note": "Published the architecture and training report for the LFM2 family.",
     "year": "2025"
    },
    {
     "title": "LFM2.5 and LFM2-24B-A2B",
     "note": "Extended the family to newer small models and a 24B mixture-of-experts model.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "cookbook",
     "url": "https://github.com/Liquid4All/cookbook",
     "note": "Examples and end-to-end apps for Liquid models."
    },
    {
     "name": "liquid-audio",
     "url": "https://github.com/Liquid4All/liquid-audio",
     "note": "Speech-to-speech audio models."
    },
    {
     "name": "antidoom",
     "url": "https://github.com/Liquid4All/antidoom",
     "note": "Targeted preference data and training."
    },
    {
     "name": "liquid-finetune",
     "url": "https://github.com/Liquid4All/liquid-finetune",
     "note": "Model customisation repo."
    }
   ],
   "hf_highlights": [
    {
     "name": "LFM2.5-Audio-1.5B-JP-GGUF",
     "url": "https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP-GGUF",
     "kind": "model",
     "note": "Top downloaded model."
    },
    {
     "name": "LFM3-Omni-2B-GGUF",
     "url": "https://huggingface.co/LiquidAI/LFM3-Omni-2B-GGUF",
     "kind": "model",
     "note": "Omni model in GGUF."
    },
    {
     "name": "nanobeir-multilingual-extended",
     "url": "https://huggingface.co/datasets/LiquidAI/nanobeir-multilingual-extended",
     "kind": "dataset",
     "note": "Multilingual retrieval evaluation data."
    }
   ],
   "papers": [
    {
     "title": "LFM2 Technical Report",
     "url": "https://arxiv.org/abs/2511.23404",
     "year": "2025"
    },
    {
     "title": "STAR: Synthesis of Tailored Architectures",
     "url": "https://arxiv.org/abs/2411.17800",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "LFM2 models",
     "url": "https://huggingface.co/LiquidAI",
     "note": "Open-weight family"
    },
    {
     "name": "cookbook",
     "url": "https://github.com/Liquid4All/cookbook",
     "note": "Examples"
    },
    {
     "name": "liquid-audio",
     "url": "https://github.com/Liquid4All/liquid-audio",
     "note": "Speech models"
    }
   ],
   "openness": "Open weights under the LFM Open License.",
   "key_people": [
    {
     "name": "Ramin Hasani",
     "role": "Co-founder and CEO",
     "url": "https://www.liquid.ai/company/about",
     "note": null
    },
    {
     "name": "Mathias Lechner",
     "role": "Co-founder and CTO",
     "url": "https://www.liquid.ai/company/about",
     "note": null
    },
    {
     "name": "Alexander Amini",
     "role": "Co-founder and Chief Scientific Officer",
     "url": "https://www.liquid.ai/company/about",
     "note": null
    },
    {
     "name": "Daniela Rus",
     "role": "Co-founder, MIT CSAIL",
     "url": "https://www.liquid.ai/company/about",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.liquid.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.liquid.ai",
    "https://huggingface.co/LiquidAI",
    "https://github.com/Liquid4All"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "liquid-ai",
   "mlat": 42.40896,
   "mlon": -71.1097,
   "logo": true
  },
  {
   "orig_id": "llm-jp",
   "name": "LLM-jp",
   "alias": null,
   "full_name": "LLM-jp",
   "type": "collective",
   "founded": 2023,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.69,
   "lon": 139.75,
   "website": "https://llm-jp.nii.ac.jp/en/",
   "huggingface": "https://huggingface.co/llm-jp",
   "github": "https://github.com/llm-jp",
   "x": "https://twitter.com/llm_jp",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Community of over 1,000 researchers and engineers coordinated by Japan's National Institute of Informatics, building Japanese-capable open LLMs.",
   "tagline": "Community of researchers organized by Japan's National Institute of Informatics, building open Japanese-capable large language models.",
   "about": [
    "LLM-jp is a researcher community organized by the Research and Development Center for Large Language Models at the National Institute of Informatics. NII established the volunteer group in May 2023, and by December 2024 it counted more than 1,900 participants. Sadao Kurohashi leads the center and the effort.",
    "It organizes working groups on corpus construction, model building, evaluation and tuning, safety, dialogue, multimodal and academic domains. The group states goals of developing open state-of-the-art Japanese models, sharing data and compute between organizations, and publishing models, tools and technical materials.",
    "The LLM-jp-3 series, trained on llm-jp-corpus v3, has been released since September 2024 in sizes from 150 million to 172 billion parameters, though access to the 172B pretrained model needs approval and redistribution is restricted."
   ],
   "mission": "LLM-jp aims to build open Japanese large language models, share knowledge among researchers and enable cross-organization collaboration on data and compute.",
   "notice": null,
   "known_for": [
    "LLM-jp-3 model series",
    "Open Japanese corpus",
    "Large cross-organization research community"
   ],
   "focus": [
    "language models",
    "Japanese",
    "datasets"
   ],
   "contributions": [
    {
     "title": "LLM-jp established",
     "note": "NII founded the volunteer research community.",
     "year": "2023"
    },
    {
     "title": "Model suite v1.0",
     "note": "First open model suite, with 13B models and fine-tuned variants, released 20 October 2023.",
     "year": "2023"
    },
    {
     "title": "Model suite v2.0 and LLM-jp-3",
     "note": "v2.0 followed on 30 April 2024, and the LLM-jp-3 series was released from September 2024.",
     "year": "2024"
    },
    {
     "title": "LLM-jp-3 MoE series",
     "note": "Extended the family with mixture-of-experts models.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "awesome-japanese-llm",
     "url": "https://github.com/llm-jp/awesome-japanese-llm",
     "note": "Overview of Japanese LLMs."
    },
    {
     "name": "llm-jp-eval",
     "url": "https://github.com/llm-jp/llm-jp-eval",
     "note": "Evaluation tool."
    },
    {
     "name": "llm-jp-sft",
     "url": "https://github.com/llm-jp/llm-jp-sft",
     "note": "SFT code."
    },
    {
     "name": "llm-jp-judge",
     "url": "https://github.com/llm-jp/llm-jp-judge",
     "note": "Automated generation evaluation."
    }
   ],
   "hf_highlights": [
    {
     "name": "llm-jp-4-33b-thinking-gguf",
     "url": "https://huggingface.co/llm-jp/llm-jp-4-33b-thinking-gguf",
     "kind": "model",
     "note": "33B thinking model."
    },
    {
     "name": "llm-jp-4.1-8b-thinking",
     "url": "https://huggingface.co/llm-jp/llm-jp-4.1-8b-thinking",
     "kind": "model",
     "note": "8B thinking model."
    },
    {
     "name": "llm-jp-3-150m",
     "url": "https://huggingface.co/llm-jp/llm-jp-3-150m",
     "kind": "model",
     "note": "Small model."
    },
    {
     "name": "llm-jp-4.1-thinking-sft-data",
     "url": "https://huggingface.co/datasets/llm-jp/llm-jp-4.1-thinking-sft-data",
     "kind": "dataset",
     "note": "SFT data."
    }
   ],
   "papers": [
    {
     "title": "LLM-jp: A Cross-organizational Project for the Research and Development of Fully Open Japanese LLMs",
     "url": "https://arxiv.org/abs/2407.03963",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "llm-jp-4-33b-thinking-gguf",
     "url": "https://huggingface.co/llm-jp/llm-jp-4-33b-thinking-gguf",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "llm-jp-4.1-8b-thinking-gguf",
     "url": "https://huggingface.co/llm-jp/llm-jp-4.1-8b-thinking-gguf",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open models, tools and corpora on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Sadao Kurohashi",
     "role": "Director of the NII Research and Development Center for Large Language Models; NII Director-General",
     "url": "https://www.nii.ac.jp/en/news/release/2024/0401.html",
     "note": "Leads LLM-jp; also Program-Specific Professor at Kyoto University."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its project website at NII.",
   "sources": [
    "https://huggingface.co/llm-jp",
    "https://llm-jp.nii.ac.jp/en/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "llm-jp",
   "logo": true
  },
  {
   "orig_id": "lmarena",
   "name": "LMArena",
   "alias": "LMSYS",
   "full_name": "LMArena (LMSYS)",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://lmarena.ai",
   "huggingface": "https://huggingface.co/lmarena-ai",
   "github": "https://github.com/lm-sys",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Berkeley-origin group behind Chatbot Arena, FastChat/Vicuna and RouteLLM; releases open evaluation data and tools.",
   "tagline": "Berkeley-origin team behind Chatbot Arena, the crowd-sourced AI model leaderboard, now operating as the company Arena.",
   "about": [
    "Chatbot Arena was released on 24 April 2023 by researchers around UC Berkeley. Its founders are Wei-Lin Chiang, Anastasios N. Angelopoulos and Ion Stoica. The site moved to its own domain, lmarena.ai, in September 2024.",
    "LMArena incorporated as an independent company in April 2025 and raised a 100 million dollar seed round that May, a 150 million dollar Series A in January 2026 and a 200 million dollar Series B in October 2026. On 28 January 2026 it rebranded as Arena. The project originated in the LMSYS group, which also produced FastChat and RouteLLM.",
    "The platform ranks models through human pairwise votes and publishes evaluation data and tools on Hugging Face and GitHub."
   ],
   "mission": "The group aims to provide transparent, community-driven evaluation of AI models using real user preferences.",
   "notice": null,
   "known_for": [
    "Chatbot Arena leaderboard",
    "Human preference evaluation",
    "FastChat and RouteLLM tools"
   ],
   "focus": [
    "evaluation",
    "language models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Chatbot Arena",
     "note": "Released a public platform ranking language models by human pairwise votes.",
     "year": "2023"
    },
    {
     "title": "Own domain",
     "note": "Moved the leaderboard to lmarena.ai.",
     "year": "2024"
    },
    {
     "title": "Company incorporated",
     "note": "LMArena became an independent company and raised a 100 million dollar seed round.",
     "year": "2025"
    },
    {
     "title": "Rebrand to Arena and Series A",
     "note": "Rebranded and raised 150 million dollars.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "FastChat",
     "url": "https://github.com/lm-sys/FastChat",
     "note": "Platform for training, serving and evaluating LLM chatbots."
    },
    {
     "name": "RouteLLM",
     "url": "https://github.com/lm-sys/RouteLLM",
     "note": "LLM router framework."
    },
    {
     "name": "llm-decontaminator",
     "url": "https://github.com/lm-sys/llm-decontaminator",
     "note": "Benchmark decontamination code."
    },
    {
     "name": "vicuna-blog-eval",
     "url": "https://github.com/lm-sys/vicuna-blog-eval",
     "note": "GPT-4 based benchmark for Vicuna."
    }
   ],
   "hf_highlights": [
    {
     "name": "leaderboard-dataset",
     "url": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "kind": "dataset",
     "note": "Leaderboard data."
    },
    {
     "name": "arena-human-preference-140k",
     "url": "https://huggingface.co/datasets/lmarena-ai/arena-human-preference-140k",
     "kind": "dataset",
     "note": "Human preference battles."
    },
    {
     "name": "VisionArena-Chat",
     "url": "https://huggingface.co/datasets/lmarena-ai/VisionArena-Chat",
     "kind": "dataset",
     "note": "Vision chat conversations."
    }
   ],
   "papers": [
    {
     "title": "Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference",
     "url": "https://arxiv.org/abs/2403.04132",
     "year": "2024"
    },
    {
     "title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena",
     "url": "https://arxiv.org/abs/2306.05685",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "FastChat",
     "url": "https://github.com/lm-sys/FastChat",
     "note": "Serving and training platform"
    },
    {
     "name": "RouteLLM",
     "url": "https://github.com/lm-sys/RouteLLM",
     "note": "Routing framework"
    }
   ],
   "openness": "Open code and preference datasets.",
   "key_people": [
    {
     "name": "Wei-Lin Chiang",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Chatbot_Arena",
     "note": null
    },
    {
     "name": "Anastasios N. Angelopoulos",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Chatbot_Arena",
     "note": null
    },
    {
     "name": "Ion Stoica",
     "role": "Co-founder; professor at UC Berkeley",
     "url": "https://en.wikipedia.org/wiki/Chatbot_Arena",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://lmarena.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://lmarena.ai",
    "https://huggingface.co/lmarena-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "lmarena",
   "mlat": 37.88862,
   "mlon": -122.46788,
   "logo": true
  },
  {
   "orig_id": "maincode",
   "name": "Maincode",
   "alias": null,
   "full_name": "Maincode",
   "type": "startup",
   "founded": null,
   "country": "Australia",
   "city": "Melbourne",
   "lat": -37.8136,
   "lon": 144.9631,
   "website": "https://www.maincode.com",
   "huggingface": "https://huggingface.co/Maincode",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "hello@maincode.com",
   "summary": "Maincode is a Melbourne AI company that describes itself as independent and runs its own hardware. It has released some models on Hugging Face, including Maincoder-1B and Matilda base models, while its main Matilda product is not open-weight.",
   "tagline": "Melbourne AI company that owns its own GPU cluster, builds the Matilda assistant and has released open Maincoder models.",
   "about": [
    "Maincode is an AI lab founded in Melbourne in late 2024. It describes itself as an independent Australian lab that runs its own hardware, including MC-2, an AMD MI355X cluster in Melbourne, with an inference engine written in Rust and a further cluster, MC-3, planned.",
    "Its product is Matilda, offered for chat, coding, desktop and API use with paid plans from 25 dollars a month. Maincode has also released Maincoder-1B, which it says scored 76 percent on HumanEval, the highest among open 1B models in its December 2025 comparison.",
    "The company's founders are not named on its homepage."
   ],
   "mission": "Maincode states that it builds AI in Australia on its own infrastructure and releases some models openly.",
   "notice": null,
   "known_for": [
    "Matilda assistant",
    "Maincoder-1B open model",
    "Owned AMD GPU cluster in Melbourne"
   ],
   "focus": [
    "language models",
    "code"
   ],
   "contributions": [
    {
     "title": "Maincode founded",
     "note": "Founded in Melbourne as an independent AI lab.",
     "year": "2024"
    },
    {
     "title": "Maincoder-1B",
     "note": "Released an Apache 2.0 code model reporting 76 percent on HumanEval.",
     "year": "2025"
    },
    {
     "title": "AMD GPU cluster investment",
     "note": "Announced a 30 million dollar investment in a Melbourne AI factory powered by AMD.",
     "year": "2025"
    },
    {
     "title": "Matilda open beta",
     "note": "Announced the open beta of its Matilda assistant on 14 July 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "Matilda-Platypus-13B-Base",
     "url": "https://huggingface.co/Maincode/Matilda-Platypus-13B-Base",
     "kind": "model",
     "note": "Matilda 13B."
    },
    {
     "name": "Maincoder-1B",
     "url": "https://huggingface.co/Maincode/Maincoder-1B",
     "kind": "model",
     "note": "Maincoder."
    },
    {
     "name": "Matilda-Platypus-1.6B-Base",
     "url": "https://huggingface.co/Maincode/Matilda-Platypus-1.6B-Base",
     "kind": "model",
     "note": "Matilda 1.6B."
    },
    {
     "name": "matilda-jev-decision-index",
     "url": "https://huggingface.co/datasets/Maincode/matilda-jev-decision-index",
     "kind": "dataset",
     "note": "Decision index."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Maincoder-1B",
     "url": "https://huggingface.co/Maincode/Maincoder-1B",
     "note": "Small code model"
    },
    {
     "name": "Matilda-Platypus-13B-Base",
     "url": "https://huggingface.co/Maincode/Matilda-Platypus-13B-Base",
     "note": "Base model"
    }
   ],
   "openness": "Selected models are open on Hugging Face; the commercial flagship is not. Borderline inclusion.",
   "key_people": [
    {
     "name": "Dave Lemphers",
     "role": "Founder and CEO",
     "url": "https://www.forbes.com.au/news/innovation/ed-cravens-multi-million-dollar-bet-on-australias-answer-to-openai/",
     "note": "Returned to Melbourne after a decade building AI startups in the United States."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.maincode.com).",
   "sources": [
    "https://www.maincode.com",
    "https://huggingface.co/Maincode"
   ],
   "confidence": "high",
   "region": "Oceania",
   "id": "maincode",
   "logo": true
  },
  {
   "orig_id": "mainlp",
   "name": "MaiNLP",
   "alias": "LMU Munich",
   "full_name": "MaiNLP (LMU Munich)",
   "type": "academic",
   "founded": null,
   "country": "Germany",
   "city": "Munich",
   "lat": 48.1351,
   "lon": 11.582,
   "website": "https://mainlp.github.io",
   "huggingface": null,
   "github": "https://github.com/mainlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Munich AI and NLP lab at LMU Munich, publishing datasets and code on human label variation, relation extraction and low-resource Germanic languages.",
   "tagline": "MaiNLP is the Munich AI and NLP lab at LMU Munich, led by Barbara Plank, working on robust and inclusive human-facing NLP.",
   "about": [
    "MaiNLP is a research lab at the Center for Information and Language Processing at LMU Munich. It is led by Barbara Plank, who holds the Chair for AI and Computational Linguistics and co-directs the center.",
    "Its research addresses robustness to language variation, fairness and human label variation, and it runs the ERC project DIALECT. The lab belongs to the ELLIS Europe NLP Program and ELLIS Unit Munich, and publishes datasets such as CrossRE and a survey of Germanic low-resource corpora on GitHub."
   ],
   "mission": "To make NLP models more robust and inclusive by studying variation in language and human labels.",
   "notice": null,
   "known_for": [
    "Human label variation research",
    "CrossRE relation extraction dataset",
    "Germanic low-resource language corpora"
   ],
   "focus": [
    "NLP"
   ],
   "contributions": [
    {
     "title": "GitHub organization opened",
     "note": "MaiNLP began publishing code on GitHub.",
     "year": "2022"
    },
    {
     "title": "Awesome human label variation",
     "note": "A curated list of datasets with un-aggregated labels accompanies an EMNLP 2022 paper.",
     "year": "2022"
    },
    {
     "title": "CrossRE",
     "note": "A cross-domain relation extraction dataset was released at EMNLP Findings.",
     "year": "2022"
    },
    {
     "title": "Germanic low-resource corpora survey",
     "note": "An overview of corpora for Germanic dialects appeared at NoDaLiDa 2023.",
     "year": "2023"
    },
    {
     "title": "Multilingual refusal and explaind",
     "note": "Code on multilingual refusal and a framework attributing model behavior was released.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "awesome-human-label-variation",
     "url": "https://github.com/mainlp/awesome-human-label-variation",
     "note": "A curated list of awesome datasets with human label variation (un-aggregated labels) in Natural Language Proce (104 stars)"
    },
    {
     "name": "CrossRE",
     "url": "https://github.com/mainlp/CrossRE",
     "note": "CrossRE: A Cross-Domain Dataset for Relation Extraction (Findings of EMNLP 2022) (49 stars)"
    },
    {
     "name": "germanic-lrl-corpora",
     "url": "https://github.com/mainlp/germanic-lrl-corpora",
     "note": "Overview of corpora/datasets for Germanic low-resource languages and dialects. Accompanies \"A Survey of Corpor (28 stars)"
    },
    {
     "name": "Multilingual-Refusal",
     "url": "https://github.com/mainlp/Multilingual-Refusal",
     "note": "Open-source repository. (19 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "awesome-human-label-variation",
     "url": "https://github.com/mainlp/awesome-human-label-variation",
     "note": "Curated datasets with un-aggregated labels"
    },
    {
     "name": "CrossRE",
     "url": "https://github.com/mainlp/CrossRE",
     "note": "Cross-domain relation extraction dataset"
    },
    {
     "name": "germanic-lrl-corpora",
     "url": "https://github.com/mainlp/germanic-lrl-corpora",
     "note": "Overview of corpora for Germanic low-resource languages"
    },
    {
     "name": "Multilingual-Refusal",
     "url": "https://github.com/mainlp/Multilingual-Refusal",
     "note": "Code on multilingual refusal in LLMs"
    }
   ],
   "openness": "Datasets and code are public on GitHub.",
   "key_people": [
    {
     "name": "Barbara Plank",
     "role": "Head of MaiNLP",
     "url": "https://github.com/bplank",
     "note": null
    },
    {
     "name": "Michael A. Hedderich",
     "role": "Junior research group leader",
     "url": "https://github.com/m-hedderich",
     "note": null
    },
    {
     "name": "Diego Frassinelli",
     "role": "Lecturer",
     "url": "https://mainlp.github.io",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://mainlp.github.io",
    "https://github.com/mainlp"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "mainlp",
   "mlat": 48.18033,
   "mlon": 11.63123,
   "logo": true
  },
  {
   "orig_id": "marin",
   "name": "Marin",
   "alias": "Open Athena",
   "full_name": "Marin (Open Athena)",
   "type": "academic",
   "founded": 2024,
   "country": "United States",
   "city": "Stanford",
   "lat": 37.4275,
   "lon": -122.1697,
   "website": "https://marin.community",
   "huggingface": "https://huggingface.co/marin-community",
   "github": "https://github.com/marin-community",
   "x": null,
   "discord": "https://discord.com/invite/J9CTk7pqcM",
   "linkedin": null,
   "contact_email": null,
   "summary": "Open lab for developing foundation models with fully open experiments; releases Marin 8B and 32B.",
   "tagline": "Open lab begun at Stanford that develops foundation models in public, releasing code, data, weights and experiment logs for Marin 8B and 32B.",
   "about": [
    "Marin is an open lab started in 2024 at Stanford University by David Hall and Percy Liang. It is primarily developed by Open Athena with community contributors, and it releases its discussions, code, model weights, data sources, successful and failed experiments and tools for learning.",
    "The lab released an 8B model in May 2025 and a 32B model in October 2025. Google's open source team reported that Marin 32B beat Olmo 2 32B on 32 of 42 tasks, and the model is released under Apache 2.0. Later work covers mixture-of-experts training, a scaling suite called Delphi, a 25-trillion-token data pipeline and post-training data.",
    "Everything is documented in public pages so others can reproduce and extend runs."
   ],
   "mission": "Marin aims to develop foundation models in the open, sharing every step from data to weights so the community can learn from and build on them.",
   "notice": null,
   "known_for": [
    "Fully open Marin 8B and 32B",
    "Open experiment record",
    "Stanford and Open Athena collaboration"
   ],
   "focus": [
    "language models",
    "reproducibility"
   ],
   "contributions": [
    {
     "title": "Marin begins",
     "note": "The project started at Stanford.",
     "year": "2024"
    },
    {
     "title": "Marin 8B",
     "note": "Released the first 8B base model.",
     "year": "2025"
    },
    {
     "title": "Marin 32B",
     "note": "Released a 32B model, which Google reports beat Olmo 2 32B on 32 of 42 tasks.",
     "year": "2025"
    },
    {
     "title": "Delphi scaling suite",
     "note": "Published a suite of scaling experiments.",
     "year": "2026"
    },
    {
     "title": "25T-token data pipeline",
     "note": "Published a description of the large data pipeline.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "marin",
     "url": "https://github.com/marin-community/marin",
     "note": "Framework for foundation model R&D."
    },
    {
     "name": "levanter",
     "url": "https://github.com/marin-community/levanter",
     "note": "Legible scalable foundation model training in JAX."
    },
    {
     "name": "haliax",
     "url": "https://github.com/marin-community/haliax",
     "note": "Named tensors for JAX."
    },
    {
     "name": "marin-experiments",
     "url": "https://github.com/marin-community/marin-experiments",
     "note": "Experiments and examples."
    }
   ],
   "hf_highlights": [
    {
     "name": "marin-8b-base",
     "url": "https://huggingface.co/marin-community/marin-8b-base",
     "kind": "model",
     "note": "Base 8B."
    },
    {
     "name": "marin-32b-base",
     "url": "https://huggingface.co/marin-community/marin-32b-base",
     "kind": "model",
     "note": "Base 32B."
    },
    {
     "name": "marin-8b-instruct",
     "url": "https://huggingface.co/marin-community/marin-8b-instruct",
     "kind": "model",
     "note": "Instruct 8B."
    },
    {
     "name": "wikipedia-markdown",
     "url": "https://huggingface.co/datasets/marin-community/wikipedia-markdown",
     "kind": "dataset",
     "note": "Markdown Wikipedia."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "marin",
     "url": "https://github.com/marin-community/marin",
     "note": "Framework"
    },
    {
     "name": "marin-32b-base",
     "url": "https://huggingface.co/marin-community/marin-32b-base",
     "note": "Model"
    }
   ],
   "openness": "Fully open code, data and logs.",
   "key_people": [
    {
     "name": "David Hall",
     "role": "Leads Marin at Open Athena",
     "url": "https://dlwh.org/",
     "note": "Created the Levanter and Haliax JAX libraries; co-founded Semantic Machines."
    },
    {
     "name": "Percy Liang",
     "role": "Co-creator of Marin; Professor, Stanford University",
     "url": "https://cs.stanford.edu/~pliang/",
     "note": "Director of the Stanford Center for Research on Foundation Models."
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://marin.community); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://marin.community",
    "https://huggingface.co/marin-community"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "marin",
   "mlat": 37.4275,
   "mlon": -122.10674,
   "logo": true
  },
  {
   "orig_id": "mbzuai",
   "name": "MBZUAI",
   "alias": null,
   "full_name": "MBZUAI",
   "type": "academic",
   "founded": 2019,
   "country": "United Arab Emirates",
   "city": "Abu Dhabi",
   "lat": 24.45,
   "lon": 54.38,
   "website": "https://mbzuai.ac.ae/",
   "huggingface": "https://huggingface.co/MBZUAI",
   "github": "https://github.com/mbzuai-oryx",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Graduate-level AI university in Abu Dhabi that publishes models and datasets on Hugging Face.",
   "tagline": "Graduate-level AI university in Abu Dhabi that runs the Institute of Foundation Models and publishes open models such as Jais and K2.",
   "about": [
    "The Mohamed bin Zayed University of Artificial Intelligence was initiated in October 2019 and licensed by the UAE Ministry of Education in April 2020. Sir J. Michael Brady served as founding interim president, and Eric Xing has been president since January 2021. Ling Shao is founding provost and Timothy Baldwin has been provost since April 2024.",
    "Its graduate programs launched in January 2021 and the first MSc cohort graduated in January 2023. By 2025 it had 84 faculty and more than 200 postdocs and researchers.",
    "The Institute of Foundation Models, launched in May 2025, coordinates the Jais Arabic models and the K2 family. In September 2026 it released K2 Horizon, six open models from 0.9B to 375B parameters, with weights, code, data and methodology."
   ],
   "mission": "The university aims to advance AI education and research, and its institute releases foundation models openly.",
   "notice": null,
   "known_for": [
    "Jais Arabic language models",
    "K2 and K2 Horizon open models",
    "AI-focused graduate university"
   ],
   "focus": [
    "research",
    "language models",
    "vision"
   ],
   "contributions": [
    {
     "title": "Initiative launched",
     "note": "The university project began.",
     "year": "2019"
    },
    {
     "title": "University license granted",
     "note": "UAE Ministry of Education licensed the university.",
     "year": "2020"
    },
    {
     "title": "First graduate programs",
     "note": "Graduate programs began.",
     "year": "2021"
    },
    {
     "title": "First MSc cohort",
     "note": "The first MSc cohort graduated.",
     "year": "2023"
    },
    {
     "title": "Institute of Foundation Models",
     "note": "Launched the institute and a Silicon Valley lab.",
     "year": "2025"
    },
    {
     "title": "K2 Horizon",
     "note": "Released six open models.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "Awesome-LLM-Post-training",
     "url": "https://github.com/mbzuai-oryx/Awesome-LLM-Post-training",
     "note": "Survey and guide on LLM post-training."
    },
    {
     "name": "Video-ChatGPT",
     "url": "https://github.com/mbzuai-oryx/Video-ChatGPT",
     "note": "Video conversation model."
    },
    {
     "name": "groundingLMM",
     "url": "https://github.com/mbzuai-oryx/groundingLMM",
     "note": "GLaMM grounded multimodal model."
    },
    {
     "name": "LLaVA-pp",
     "url": "https://github.com/mbzuai-oryx/LLaVA-pp",
     "note": "LLaVA with Phi-3 and Llama-3."
    }
   ],
   "hf_highlights": [
    {
     "name": "MBZUAI/AIN",
     "url": "https://huggingface.co/MBZUAI/AIN",
     "kind": "model",
     "note": "Arabic-English multimodal model."
    },
    {
     "name": "MBZUAI/GeoPixel-7B",
     "url": "https://huggingface.co/MBZUAI/GeoPixel-7B",
     "kind": "model",
     "note": "Remote-sensing grounding model."
    },
    {
     "name": "MBZUAI/LaMini-GPT-124M",
     "url": "https://huggingface.co/MBZUAI/LaMini-GPT-124M",
     "kind": "model",
     "note": "Distilled small model."
    },
    {
     "name": "MBZUAI/Omni-Sets",
     "url": "https://huggingface.co/datasets/MBZUAI/Omni-Sets",
     "kind": "dataset",
     "note": "Dataset."
    }
   ],
   "papers": [
    {
     "title": "Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models",
     "url": "https://arxiv.org/abs/2308.16149",
     "year": "2023"
    },
    {
     "title": "GLaMM: Pixel Grounding Large Multimodal Model",
     "url": "https://arxiv.org/abs/2311.03356",
     "year": "2023"
    },
    {
     "title": "Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models",
     "url": "https://arxiv.org/abs/2306.05424",
     "year": "2023"
    },
    {
     "title": "LaMini-LM: A Diverse Herd of Distilled Models from Large-Scale Instructions",
     "url": "https://arxiv.org/abs/2304.14402",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "TerraFM",
     "url": "https://huggingface.co/MBZUAI/TerraFM",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "MedChain-R1",
     "url": "https://huggingface.co/MBZUAI/MedChain-R1",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Models on Hugging Face; Arabic and multilingual releases.",
   "key_people": [
    {
     "name": "Eric Xing",
     "role": "President",
     "url": "https://en.wikipedia.org/wiki/Mohamed_bin_Zayed_University_of_Artificial_Intelligence",
     "note": null
    },
    {
     "name": "Timothy Baldwin",
     "role": "Provost",
     "url": "https://en.wikipedia.org/wiki/Mohamed_bin_Zayed_University_of_Artificial_Intelligence",
     "note": null
    },
    {
     "name": "Ling Shao",
     "role": "Founding provost and executive vice president",
     "url": "https://en.wikipedia.org/wiki/Mohamed_bin_Zayed_University_of_Artificial_Intelligence",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact routes are on mbzuai.ac.ae; individual research groups keep their own GitHub pages.",
   "sources": [
    "https://huggingface.co/MBZUAI",
    "https://mbzuai.ac.ae/"
   ],
   "confidence": "medium",
   "region": "Middle East",
   "id": "mbzuai",
   "mlat": 24.47165,
   "mlon": 54.42119,
   "logo": true
  },
  {
   "orig_id": "dbmdz",
   "name": "MDZ Digital Library team",
   "alias": "Bavarian State Library",
   "full_name": "MDZ Digital Library team (Bavarian State Library)",
   "type": "institute",
   "founded": null,
   "country": "Germany",
   "city": "Munich",
   "lat": 48.1351,
   "lon": 11.582,
   "website": "https://github.com/dbmdz",
   "huggingface": "https://huggingface.co/dbmdz",
   "github": "https://github.com/dbmdz",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Digital library team of the Bavarian State Library known for the widely used dbmdz German, Turkish and Italian BERT and ELECTRA models.",
   "tagline": "The MDZ Digital Library team at the Bavarian State Library publishes the widely used dbmdz German, Italian and Turkish BERT models.",
   "about": [
    "The Munich Digitization Center of the Bavarian State Library runs the dbmdz GitHub organization, which began in 2016 and mixes digital library software with open language models. Its berts repository, started in September 2019, documents BERT, DistilBERT, ELECTRA, GPT-2 and ConvBERT models.",
    "Models trained on Europeana newspapers cover German, French and historic multilingual text, with German BERT released in February 2020. All models are MIT-licensed on Hugging Face under the dbmdz organization."
   ],
   "mission": "To make digital library collections and language models for historical and modern text openly usable.",
   "notice": null,
   "known_for": [
    "dbmdz German BERT and ELECTRA models",
    "Europeana newspaper historic models",
    "IIIF and OCR open-source tools"
   ],
   "focus": [
    "language models",
    "digital libraries",
    "OCR"
   ],
   "contributions": [
    {
     "title": "berts repository",
     "note": "The dbmdz models repository was opened.",
     "year": "2019"
    },
    {
     "title": "Europeana German BERT",
     "note": "Cased and uncased BERT models trained on Europeana newspapers were released.",
     "year": "2020"
    },
    {
     "title": "German GPT-2 and Europeana French",
     "note": "A German GPT-2 and French BERT and ELECTRA models were released.",
     "year": "2020"
    },
    {
     "title": "Historic language models",
     "note": "Multilingual and monolingual historic models were published.",
     "year": "2021"
    },
    {
     "title": "Historic Dutch BERT",
     "note": "A model trained on Delpher newspapers from 1618 to 1879 was released.",
     "year": "2021"
    }
   ],
   "github_highlights": [
    {
     "name": "berts",
     "url": "https://github.com/dbmdz/berts",
     "note": "DBMDZ BERT, DistilBERT, ELECTRA, GPT-2 and ConvBERT models (157 stars)"
    },
    {
     "name": "solr-ocrhighlighting",
     "url": "https://github.com/dbmdz/solr-ocrhighlighting",
     "note": "Highlighting various OCR formats directly in Solr (88 stars)"
    },
    {
     "name": "deep-eos",
     "url": "https://github.com/dbmdz/deep-eos",
     "note": "General-Purpose Neural Networks for Sentence Boundary Detection (74 stars)"
    },
    {
     "name": "mirador-textoverlay",
     "url": "https://github.com/dbmdz/mirador-textoverlay",
     "note": "Text Overlay plugin for Mirador 4 (63 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "dbmdz/bert-base-turkish-cased",
     "url": "https://huggingface.co/dbmdz/bert-base-turkish-cased",
     "kind": "model",
     "note": "Model with 192 likes on the Hub."
    },
    {
     "name": "dbmdz/bert-large-cased-finetuned-conll03-english",
     "url": "https://huggingface.co/dbmdz/bert-large-cased-finetuned-conll03-english",
     "kind": "model",
     "note": "Model with 98 likes on the Hub."
    },
    {
     "name": "dbmdz/detectron2-model-demo",
     "url": "https://huggingface.co/spaces/dbmdz/detectron2-model-demo",
     "kind": "space",
     "note": "Space with 5 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "dbmdz BERT models",
     "url": "https://github.com/dbmdz/berts",
     "note": "DBMDZ BERT, DistilBERT, ELECTRA, GPT-2 and ConvBERT models"
    },
    {
     "name": "BERTurk",
     "url": "https://huggingface.co/dbmdz/bert-base-turkish-cased",
     "note": "Turkish BERT model"
    },
    {
     "name": "deep-eos",
     "url": "https://github.com/dbmdz/deep-eos",
     "note": "Sentence boundary detection networks"
    },
    {
     "name": "solr-ocrhighlighting",
     "url": "https://github.com/dbmdz/solr-ocrhighlighting",
     "note": "OCR highlighting plugin for Solr"
    }
   ],
   "openness": "All checked models and tools are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Stefan Schweter",
     "role": "Model author",
     "url": "https://github.com/stefan-it",
     "note": null
    },
    {
     "name": "Johannes Baiter",
     "role": "Developer",
     "url": "https://github.com/jbaiter",
     "note": null
    },
    {
     "name": "Marcus Bitzl",
     "role": "Developer",
     "url": "https://github.com/bitzl",
     "note": null
    },
    {
     "name": "Katharina Schmid",
     "role": "Member",
     "url": "https://huggingface.co/schmika",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the library team site.",
   "sources": [
    "https://github.com/dbmdz",
    "https://huggingface.co/dbmdz",
    "https://github.com/dbmdz"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "mdz-digital-library-team",
   "mlat": 48.18033,
   "mlon": 11.53277,
   "logo": true
  },
  {
   "orig_id": "mesolitica",
   "name": "Mesolitica",
   "alias": "MaLLaM",
   "full_name": "Mesolitica (MaLLaM)",
   "type": "startup",
   "founded": null,
   "country": "Malaysia",
   "city": "Kuala Lumpur",
   "lat": 3.14,
   "lon": 101.69,
   "website": "https://mesolitica.com/",
   "huggingface": "https://huggingface.co/mesolitica",
   "github": "https://github.com/mesolitica",
   "x": "https://twitter.com/mesolitica",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Malaysian multimodal AI lab releasing Malay translation, TTS and language models.",
   "tagline": "Malaysian AI lab building open Malay language models, speech systems and datasets, including MaLLaM and Malaysian Mistral.",
   "about": [
    "Mesolitica was founded in Kuala Lumpur by Khalil Nooh, chief executive, and Husein Zolkepli, chief technology officer. The team traces the work behind MaLLaM to January 2019, and its GitHub organization dates from February 2019.",
    "It launched MaLLaM, a Malaysian language model trained on nearly 200 billion tokens of Malay content, in December 2023. In January 2024 it published papers on MaLLaM and on Malaysian Mistral, a Mistral 7B model finetuned on Malaysian data, and released models on Hugging Face. It also works on speech systems and voice agents.",
    "The company says it runs operations within Malaysia and open-sources its datasets on GitHub while allowing owners to request removal of their content."
   ],
   "mission": "Mesolitica aims to bring AI to Bahasa Melayu speakers and promote AI sovereignty by working entirely within Malaysia.",
   "notice": null,
   "known_for": [
    "MaLLaM Malaysian language model",
    "Malaysian Mistral",
    "Open Malay datasets"
   ],
   "focus": [
    "Malay",
    "speech",
    "translation"
   ],
   "contributions": [
    {
     "title": "Work begins",
     "note": "The effort behind MaLLaM began.",
     "year": "2019"
    },
    {
     "title": "MaLLaM",
     "note": "Launched a Malaysian LLM for businesses.",
     "year": "2023"
    },
    {
     "title": "Malaysian Mistral and MaLLaM papers",
     "note": "Published papers and released models.",
     "year": "2024"
    },
    {
     "title": "Build on AWS",
     "note": "Reported training a Malaysian LLM on AWS.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "NLP-Models-Tensorflow",
     "url": "https://github.com/mesolitica/NLP-Models-Tensorflow",
     "note": "Collection of NLP models in TensorFlow."
    },
    {
     "name": "vllm-whisper",
     "url": "https://github.com/mesolitica/vllm-whisper",
     "note": "Whisper serving engine."
    },
    {
     "name": "llm-embedding",
     "url": "https://github.com/mesolitica/llm-embedding",
     "note": "Malaysian embedding fine-tuning."
    },
    {
     "name": "multimodal-LLM",
     "url": "https://github.com/mesolitica/multimodal-LLM",
     "note": "Image, audio and text LLM."
    }
   ],
   "hf_highlights": [
    {
     "name": "mesolitica/wav2vec2-xls-r-300m-mixed",
     "url": "https://huggingface.co/mesolitica/wav2vec2-xls-r-300m-mixed",
     "kind": "model",
     "note": "Most downloaded."
    },
    {
     "name": "mesolitica/llama2-embedding-1b-8k",
     "url": "https://huggingface.co/mesolitica/llama2-embedding-1b-8k",
     "kind": "model",
     "note": "Embedding model."
    },
    {
     "name": "mesolitica/Malaysian-TTS-0.6B-v1",
     "url": "https://huggingface.co/mesolitica/Malaysian-TTS-0.6B-v1",
     "kind": "model",
     "note": "TTS."
    },
    {
     "name": "mesolitica/Malaysian-Emilia",
     "url": "https://huggingface.co/datasets/mesolitica/Malaysian-Emilia",
     "kind": "dataset",
     "note": "Speech dataset."
    }
   ],
   "papers": [
    {
     "title": "MaLLaM -- Malaysia Large Language Model",
     "url": "https://arxiv.org/abs/2401.14680",
     "year": "2024"
    },
    {
     "title": "Large Malaysian Language Model Based on Mistral for Enhanced Local Language Understanding",
     "url": "https://arxiv.org/abs/2401.13565",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Malaysian-TTS-4B-v0.1",
     "url": "https://huggingface.co/mesolitica/Malaysian-TTS-4B-v0.1",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "nanot5-base-malaysian-translation-v2",
     "url": "https://huggingface.co/mesolitica/nanot5-base-malaysian-translation-v2",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open weights.",
   "key_people": [
    {
     "name": "Khalil Nooh",
     "role": "Co-founder and CEO",
     "url": "https://www.tatlerasia.com/power-purpose/innovation/mesolitica-ai-mallam-llm-khalil-nooh",
     "note": null
    },
    {
     "name": "Husein Zolkepli",
     "role": "Co-founder and CTO",
     "url": "https://huggingface.co/huseinzol05",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "mesolitica.com is the main site; GitHub issues are the open channel.",
   "sources": [
    "https://huggingface.co/mesolitica",
    "https://mesolitica.com/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "mesolitica",
   "logo": true
  },
  {
   "orig_id": "metr",
   "name": "METR",
   "alias": null,
   "full_name": "METR",
   "type": "nonprofit",
   "founded": 2022,
   "country": "United States",
   "city": "Berkeley",
   "lat": 37.8715,
   "lon": -122.273,
   "website": "https://metr.org",
   "huggingface": null,
   "github": "https://github.com/METR",
   "x": "https://x.com/METR_Evals",
   "discord": null,
   "linkedin": null,
   "contact_email": "info@metr.org",
   "summary": "Nonprofit that evaluates frontier AI models for autonomous capabilities and publishes evaluation code and datasets.",
   "tagline": "A Berkeley nonprofit that measures what frontier AI agents can do on long, autonomous tasks and publishes the evidence for governments, labs and the public.",
   "about": [
    "METR, short for Model Evaluation and Threat Research, is a nonprofit based in Berkeley, California. It began as ARC Evals, the evaluation arm of the Alignment Research Center, and was spun out as an independent organization in December 2023 under founder and CEO Beth Barnes.",
    "The group builds task suites and measurement methods for autonomous AI agents, and publishes the results. Its best-known metric, the task-completion time horizon, expresses how long a task, measured by human expert completion time, an agent can finish at a given success rate. It also runs applied assessments with AI developers, including pre-deployment evaluations and incident investigations.",
    "METR states that it is funded by donations and has not accepted funding from AI companies, although developers provide access and tokens for evaluations. It has worked with several frontier developers and provides technical assistance to the European AI Office."
   ],
   "mission": "METR aims to discover and publish evidence on the capabilities of advanced AI systems, the risks they pose and the effectiveness of mitigations, so that the world can avoid catastrophic mistakes.",
   "notice": null,
   "known_for": [
    "Task-completion time horizon metric for AI agents",
    "Independent pre-deployment evaluations of frontier models",
    "RE-Bench and long-horizon agent task suites",
    "Open evaluation code and datasets"
   ],
   "focus": [
    "evaluation",
    "safety",
    "agents",
    "time horizons"
   ],
   "contributions": [
    {
     "title": "Spin-out from ARC Evals",
     "note": "ARC Evals became the independent nonprofit METR in December 2023.",
     "year": "2023"
    },
    {
     "title": "Time horizon analysis",
     "note": "A March 2025 paper reported that the length of software tasks leading AI models could complete doubled roughly every seven months between 2019 and 2024.",
     "year": "2025"
    },
    {
     "title": "Time Horizon 1.1",
     "note": "A revised time horizon model with more tasks and new evaluation infrastructure, released in January 2026.",
     "year": "2026"
    },
    {
     "title": "MirrorCode results",
     "note": "Early results showing agents completing weeks-long coding tasks, including reimplementing a 16,000-line codebase.",
     "year": "2026"
    },
    {
     "title": "Frontier Risk Report pilot",
     "note": "A pilot assessment of rogue deployment risk inside frontier AI developers, with Anthropic, Google, Meta and OpenAI taking part.",
     "year": "2026"
    },
    {
     "title": "Agent incident investigation",
     "note": "A report on agent behavior in an incident involving OpenAI agents and Hugging Face, published in August 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "eval-analysis-public",
     "url": "https://github.com/METR/eval-analysis-public",
     "note": "Public eval data pipeline"
    },
    {
     "name": "task-standard",
     "url": "https://github.com/METR/task-standard",
     "note": "METR task standard"
    },
    {
     "name": "RE-Bench",
     "url": "https://github.com/METR/RE-Bench",
     "note": "AI R&D evaluation tasks"
    },
    {
     "name": "vivaria",
     "url": "https://github.com/METR/vivaria",
     "note": "Tool for running evaluations"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Measuring AI Ability to Complete Long Software Tasks",
     "url": "https://arxiv.org/abs/2503.14499",
     "year": "2025"
    },
    {
     "title": "RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts",
     "url": "https://arxiv.org/abs/2411.15114",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "eval-analysis-public",
     "url": "https://github.com/METR/eval-analysis-public",
     "note": "Public eval data pipeline"
    },
    {
     "name": "task-standard",
     "url": "https://github.com/METR/task-standard",
     "note": "METR task standard"
    },
    {
     "name": "RE-Bench",
     "url": "https://github.com/METR/RE-Bench",
     "note": "AI R&D evaluation tasks"
    }
   ],
   "openness": "MIT-licensed task standard and tools; evaluation analysis data on GitHub.",
   "key_people": [
    {
     "name": "Beth Barnes",
     "role": "Founder and CEO",
     "url": "https://metr.org/about",
     "note": null
    },
    {
     "name": "Chris Painter",
     "role": "President",
     "url": "https://metr.org/about",
     "note": null
    },
    {
     "name": "Hjalmar Wijk",
     "role": "Chief Scientist",
     "url": "https://metr.org/about",
     "note": null
    },
    {
     "name": "Nate Rush",
     "role": "CTO",
     "url": "https://metr.org/about",
     "note": null
    },
    {
     "name": "Ajeya Cotra",
     "role": "Technical Staff",
     "url": "https://metr.org/about",
     "note": null
    },
    {
     "name": "Ryan Greenblatt",
     "role": "Technical Staff",
     "url": "https://metr.org/about",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on website.",
   "sources": [
    "https://metr.org",
    "https://github.com/METR"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "metr",
   "mlat": 37.91673,
   "mlon": -122.31462,
   "logo": true
  },
  {
   "orig_id": "mila",
   "name": "Mila - Quebec AI Institute",
   "alias": null,
   "full_name": "Mila - Quebec AI Institute",
   "type": "institute",
   "founded": 1993,
   "country": "Canada",
   "city": "Montreal",
   "lat": 45.5017,
   "lon": -73.5673,
   "website": "https://mila.quebec",
   "huggingface": null,
   "github": "https://github.com/mila-iqia",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@mila.quebec",
   "summary": "Research institute in Montreal founded around Yoshua Bengio's lab; its groups release open research code such as BabyAI and Theano-era frameworks.",
   "tagline": "Montreal's AI research institute, founded around Yoshua Bengio's lab, whose members built Theano, SpeechBrain and other widely used open-source tools.",
   "about": [
    "Mila, the Quebec AI Institute, traces its origin to the lab Yoshua Bengio founded at the Université de Montréal in 1993. In 2017 it grew into a collaboration between the Université de Montréal and McGill University with Polytechnique Montréal and HEC Montréal, and it became a non-profit in 2018.",
    "The institute brings together academic researchers and students working on deep learning, reinforcement learning, health, climate, responsible AI and related areas. It reports a community of more than 1,400 people and runs training programs, an AI policy fellowship and a venture program alongside research.",
    "Mila lists 34 open-source projects, including Theano, an early deep learning framework that originated there, SpeechBrain, TorchDrug, BabyAI and Orion."
   ],
   "mission": "Mila aims to be a global hub for scientific advances that drive AI innovation for everyone's benefit, built on collaboration, excellence and ethics.",
   "notice": null,
   "known_for": [
    "Origin of the Theano deep learning framework",
    "Yoshua Bengio's founding lab",
    "Open-source tools such as SpeechBrain and TorchDrug",
    "One of Canada's three national AI institutes"
   ],
   "focus": [
    "deep learning",
    "reinforcement learning",
    "research"
   ],
   "contributions": [
    {
     "title": "Founding of the lab",
     "note": "Yoshua Bengio founded the Montreal lab that later became Mila.",
     "year": "1993"
    },
    {
     "title": "Mila as a university partnership",
     "note": "The institute expanded through a collaboration of Université de Montréal and McGill University with Polytechnique Montréal and HEC Montréal.",
     "year": "2017"
    },
    {
     "title": "End of Theano development",
     "note": "Development of Theano, an early deep learning framework from Mila and the Université de Montréal, was completed.",
     "year": "2017"
    },
    {
     "title": "Non-profit status",
     "note": "Mila became a non-profit organization.",
     "year": "2018"
    },
    {
     "title": "Move to new premises",
     "note": "Mila moved into its current building in Montreal's Mile-Ex neighbourhood.",
     "year": "2019"
    }
   ],
   "github_highlights": [
    {
     "name": "blocks",
     "url": "https://github.com/mila-iqia/blocks",
     "note": "Theano framework for building and training neural networks."
    },
    {
     "name": "welcome_tutorials",
     "url": "https://github.com/mila-iqia/welcome_tutorials",
     "note": "Tutorials for welcoming new Mila students."
    },
    {
     "name": "fuel",
     "url": "https://github.com/mila-iqia/fuel",
     "note": "Data pipeline framework for ML."
    },
    {
     "name": "babyai",
     "url": "https://github.com/mila-iqia/babyai",
     "note": "Testbed for training agents to understand language."
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Neural Machine Translation by Jointly Learning to Align and Translate",
     "url": "https://arxiv.org/abs/1409.0473",
     "year": "2014"
    },
    {
     "title": "GFlowNet Foundations",
     "url": "https://arxiv.org/abs/2111.09266",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "BabyAI",
     "url": "https://github.com/mila-iqia/babyai",
     "note": "Grounded-language testbed"
    },
    {
     "name": "blocks",
     "url": "https://github.com/mila-iqia/blocks",
     "note": "Theano framework"
    }
   ],
   "openness": "Open research code; weights vary by project.",
   "key_people": [
    {
     "name": "Yoshua Bengio",
     "role": "Founder",
     "url": "https://mila.quebec/en/directory",
     "note": null
    },
    {
     "name": "Hugo Larochelle",
     "role": "Scientific Director",
     "url": "https://mila.quebec/en/directory/hugo-larochelle",
     "note": null
    },
    {
     "name": "Valérie Pisano",
     "role": "President and CEO",
     "url": "https://mila.quebec/en/directory/valerie-pisano",
     "note": null
    },
    {
     "name": "Stéphane Létourneau",
     "role": "Executive Vice President",
     "url": "https://mila.quebec/en/directory/stephane-letourneau",
     "note": null
    },
    {
     "name": "Laurence Beaulieu",
     "role": "Vice President, Academic Affairs",
     "url": "https://mila.quebec/en/directory/laurence-beaulieu",
     "note": null
    },
    {
     "name": "Frédéric Chanay-Savoyen",
     "role": "Vice-President, AI and Technology Solutions",
     "url": "https://mila.quebec/en/directory/frederic-chanay-savoyen",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://mila.quebec); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://mila.quebec",
    "https://github.com/mila-iqia"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "mila-quebec-ai-institute",
   "logo": true
  },
  {
   "orig_id": "minimax",
   "name": "MiniMax",
   "alias": null,
   "full_name": "MiniMax",
   "type": "startup",
   "founded": 2021,
   "country": "China",
   "city": "Shanghai",
   "lat": 31.23,
   "lon": 121.47,
   "website": "https://www.minimax.io",
   "huggingface": "https://huggingface.co/MiniMaxAI",
   "github": "https://github.com/MiniMax-AI",
   "x": "https://twitter.com/MiniMax_AI",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Shanghai AI company publishing the MiniMax text and image model families as open weights.",
   "tagline": "A Shanghai AI company building multimodal foundation models, with the MiniMax-M series of language models released as open weights on Hugging Face.",
   "about": [
    "MiniMax was founded in December 2021 in Shanghai by Yan Junjie, Yun Yeyi and Zhou Yucong, who came from computer vision research at SenseTime. It began research in January 2022 and released its first text model, abab1, in April 2022. The company listed on the Hong Kong Stock Exchange in January 2026.",
    "It develops general-purpose models across text, speech, music, image and video, and serves consumer products such as the Hailuo video service and the Talkie character app alongside an API platform. Its company page reports more than 300 million users in over 230 countries and regions.",
    "The MiniMax-M series of language models is published on Hugging Face under the MiniMaxAI organization, with recent releases including MiniMax-M3, M3.1-Flash and the H3 image models. MiniMax is a commercial, publicly listed company that releases some weights openly."
   ],
   "mission": "MiniMax describes its mission as Intelligence with Everyone, building its own foundation models toward artificial general intelligence.",
   "notice": null,
   "known_for": [
    "MiniMax-M series open-weight language models",
    "Hailuo video generation",
    "Talkie character chat app",
    "Long-context MiniMax-01 models"
   ],
   "focus": [
    "language models",
    "multimodal",
    "image generation",
    "Chinese",
    "English"
   ],
   "contributions": [
    {
     "title": "abab1 text model",
     "note": "MiniMax launched abab1, its first text model, in April 2022.",
     "year": "2022"
    },
    {
     "title": "Hailuo AI",
     "note": "MiniMax launched Hailuo AI in March 2024.",
     "year": "2024"
    },
    {
     "title": "MiniMax-01",
     "note": "The company unveiled the MiniMax-Text-01 and MiniMax-VL-01 model line in January 2025.",
     "year": "2025"
    },
    {
     "title": "MiniMax-M1",
     "note": "MiniMax released the M1 language model in June 2025.",
     "year": "2025"
    },
    {
     "title": "Hong Kong listing",
     "note": "MiniMax listed on the Hong Kong Stock Exchange on 9 January 2026.",
     "year": "2026"
    },
    {
     "title": "MiniMax-M3",
     "note": "MiniMax-M3 weights appeared on Hugging Face in June 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "skills",
     "url": "https://github.com/MiniMax-AI/skills",
     "note": "Agent skills collection, the org's most starred repo."
    },
    {
     "name": "MiniMax-H3",
     "url": "https://github.com/MiniMax-AI/MiniMax-H3",
     "note": "Repository for the H3 model."
    },
    {
     "name": "MiniMax-01",
     "url": "https://github.com/MiniMax-AI/MiniMax-01",
     "note": "Official repo for MiniMax-Text-01 and VL-01."
    },
    {
     "name": "MiniMax-M1",
     "url": "https://github.com/MiniMax-AI/MiniMax-M1",
     "note": "Open-weight hybrid-attention reasoning model."
    }
   ],
   "hf_highlights": [
    {
     "name": "MiniMax-H3",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3",
     "kind": "model",
     "note": "Most downloaded model in the org."
    },
    {
     "name": "MiniMax-M2.5",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5",
     "kind": "model",
     "note": "M2.5 agentic model."
    },
    {
     "name": "MiniMax-H3-Image",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3-Image",
     "kind": "model",
     "note": "Image model."
    },
    {
     "name": "SynLogic",
     "url": "https://huggingface.co/datasets/MiniMaxAI/SynLogic",
     "kind": "dataset",
     "note": "Synthetic logic reasoning data."
    }
   ],
   "papers": [
    {
     "title": "MiniMax-01: Scaling Foundation Models with Lightning Attention",
     "url": "https://arxiv.org/abs/2501.08313",
     "year": "2025"
    },
    {
     "title": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
     "url": "https://arxiv.org/abs/2506.13585",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "MiniMax-H3",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3",
     "note": "MiniMax H3"
    },
    {
     "name": "MiniMax-M2.5",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5",
     "note": "MiniMax M2.5"
    },
    {
     "name": "MiniMax-M2.7",
     "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7",
     "note": "MiniMax M2.7"
    }
   ],
   "openness": "Open weights on Hugging Face; licence per model card.",
   "key_people": [
    {
     "name": "Yan Junjie",
     "role": "Co-founder and CEO",
     "url": "https://en.wikipedia.org/wiki/MiniMax_Group",
     "note": null
    },
    {
     "name": "Yun Yeyi",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/MiniMax_Group",
     "note": null
    },
    {
     "name": "Zhou Yucong",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/MiniMax_Group",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the official site, GitHub issues or Hugging Face discussions.",
   "sources": [
    "https://huggingface.co/MiniMaxAI",
    "https://www.minimax.io"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "minimax",
   "mlat": 31.18,
   "mlon": 121.47,
   "logo": true
  },
  {
   "orig_id": "mistral-ai",
   "name": "Mistral AI",
   "alias": null,
   "full_name": "Mistral AI",
   "type": "startup",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://mistral.ai",
   "huggingface": "https://huggingface.co/mistralai",
   "github": "https://github.com/mistralai",
   "x": "https://x.com/mistralai",
   "discord": "https://discord.gg/mistralai",
   "linkedin": "https://www.linkedin.com/company/mistralai",
   "contact_email": "press@mistral.ai",
   "summary": "Paris-based company building frontier language, vision and speech models. It publishes many of its models as open weights alongside a commercial API.",
   "tagline": "A Paris AI company that pairs open-weight models under permissive licenses with a commercial platform, positioning itself as Europe's frontier lab.",
   "about": [
    "Mistral AI was founded in April 2023 in Paris by Arthur Mensch, Guillaume Lample and Timothée Lacroix, who had met at École polytechnique and worked at Google DeepMind and Meta. Its first model, Mistral 7B, was released on 27 September 2023 under the Apache 2.0 license.",
    "The company builds language, coding, speech, document and forecasting models, and sells them through its Studio and Vibe products and through sovereign deployments for enterprises and governments. Recent open-weight releases include Mistral Small 4 under Apache 2.0, Voxtral TTS and the one-trillion-parameter Mistral Large 4.",
    "Mistral is venture-funded and commercial. ASML took an 11 percent shareholding in 2025, and the firm is often cited as a leading beneficiary of European digital sovereignty policy."
   ],
   "mission": "Mistral says its goal is to put frontier AI in everyone's hands, keeping it open and giving organizations control and ownership of their AI systems.",
   "notice": null,
   "known_for": [
    "Mistral 7B and Mixtral open-weight models",
    "Apache 2.0 licensed releases",
    "Europe's largest AI startup",
    "Sovereign AI deployments for enterprises and governments"
   ],
   "focus": [
    "language models",
    "multimodal",
    "speech",
    "code"
   ],
   "contributions": [
    {
     "title": "Founding in Paris",
     "note": "Mensch, Lample and Lacroix established the company in April 2023.",
     "year": "2023"
    },
    {
     "title": "Mistral 7B",
     "note": "A 7.3-billion-parameter model released under Apache 2.0 that outperformed Llama 2 13B on the benchmarks Mistral reported.",
     "year": "2023"
    },
    {
     "title": "Magistral Small",
     "note": "The first open reasoning model from Mistral, released in June 2025.",
     "year": "2025"
    },
    {
     "title": "Mistral Large 3 and Ministral 3",
     "note": "A 675-billion-parameter sparse mixture-of-experts model and three small dense models released in December 2025.",
     "year": "2025"
    },
    {
     "title": "Mistral Small 4",
     "note": "A multimodal hybrid model with a 256,000-token context window released under Apache 2.0 in March 2026.",
     "year": "2026"
    },
    {
     "title": "Mistral Large 4",
     "note": "A one-trillion-parameter mixture-of-experts model whose weights were released on 29 October 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "mistral-inference",
     "url": "https://github.com/mistralai/mistral-inference",
     "note": "Official inference library for Mistral models (10,814 stars)"
    },
    {
     "name": "mistral-vibe",
     "url": "https://github.com/mistralai/mistral-vibe",
     "note": "Minimal CLI coding agent by Mistral (5,401 stars)"
    },
    {
     "name": "mistral-finetune",
     "url": "https://github.com/mistralai/mistral-finetune",
     "note": "Open-source repository (3,089 stars)"
    },
    {
     "name": "cookbook",
     "url": "https://github.com/mistralai/cookbook",
     "note": "Open-source repository (2,367 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "mistralai/Mistral-7B-Instruct-v0.3",
     "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3",
     "kind": "model",
     "note": "Model: 1,815,007 downloads, 3843 likes."
    },
    {
     "name": "mistralai/Voxtral-Mini-4B-Realtime-2602",
     "url": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602",
     "kind": "model",
     "note": "Model: 1,546,100 downloads, 1062 likes."
    },
    {
     "name": "mistralai/mmlu_speech",
     "url": "https://huggingface.co/datasets/mistralai/mmlu_speech",
     "kind": "dataset",
     "note": "Dataset: 483 downloads, 20 likes."
    },
    {
     "name": "mistralai/voxtral-tts-demo",
     "url": "https://huggingface.co/spaces/mistralai/voxtral-tts-demo",
     "kind": "space",
     "note": "Space: 225 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Mistral 7B",
     "url": "https://arxiv.org/abs/2310.06825",
     "year": "2023"
    },
    {
     "title": "Mixtral of Experts",
     "url": "https://arxiv.org/abs/2401.04088",
     "year": "2024"
    },
    {
     "title": "Pixtral 12B",
     "url": "https://arxiv.org/abs/2410.07073",
     "year": "2024"
    },
    {
     "title": "Voxtral",
     "url": "https://arxiv.org/abs/2507.13264",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Mistral 7B",
     "url": "https://huggingface.co/mistralai/Mistral-7B-v0.1",
     "note": "Base 7B model released as open weights"
    },
    {
     "name": "Mixtral 8x7B Instruct",
     "url": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1",
     "note": "Sparse mixture-of-experts model"
    },
    {
     "name": "mistral-inference",
     "url": "https://github.com/mistralai/mistral-inference",
     "note": "Official inference library"
    }
   ],
   "openness": "Many models are released as open weights with Apache-2.0 licences, plus open inference and finetuning code; some flagship models are API-only.",
   "key_people": [
    {
     "name": "Arthur Mensch",
     "role": "Co-founder and CEO",
     "url": "https://mistral.ai/company",
     "note": null
    },
    {
     "name": "Guillaume Lample",
     "role": "Co-founder and Chief Science Officer",
     "url": "https://mistral.ai/company",
     "note": null
    },
    {
     "name": "Timothée Lacroix",
     "role": "Co-founder and CTO",
     "url": "https://mistral.ai/company",
     "note": null
    }
   ],
   "discord_note": "Official Discord community for developers using Mistral models and API.",
   "contact_note": "Press address printed on the website; Discord community for developers.",
   "sources": [
    "https://mistral.ai",
    "https://huggingface.co/mistralai",
    "https://github.com/mistralai",
    "https://arxiv.org/abs/2310.06825"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "mistral-ai",
   "mlat": 48.94424,
   "mlon": 2.26973,
   "logo": true
  },
  {
   "orig_id": "mixedbread",
   "name": "Mixedbread",
   "alias": null,
   "full_name": "Mixedbread",
   "type": "startup",
   "founded": 2023,
   "country": "Germany",
   "city": "Berlin",
   "lat": 52.52,
   "lon": 13.405,
   "website": "https://www.mixedbread.com",
   "huggingface": "https://huggingface.co/mixedbread-ai",
   "github": "https://github.com/mixedbread-ai",
   "x": "https://x.com/mixedbreadai",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/100432189",
   "contact_email": null,
   "summary": "Berlin startup building search and retrieval models. It publishes open embedding and reranking models alongside a hosted product.",
   "tagline": "Applied research lab building retrieval and reranking models, with open embeddings downloaded more than 50 million times.",
   "about": [
    "Mixedbread is an applied research lab focused on retrieval. It builds embedding, reranking and search systems that give AI applications the right context at the right time.",
    "The team combines deep learning research with high-performance systems engineering and offers its models through a hosted search product while publishing open-source models on Hugging Face. The company states that those open models have been downloaded more than 50 million times.",
    "Recent work covers listwise rerankers, quantization for late-interaction retrieval and ingestion of very large files."
   ],
   "mission": "Mixedbread aims to build memory for AI, so that intelligent systems have the context they need to understand and act on the world.",
   "notice": null,
   "known_for": [
    "Open embedding and reranking models",
    "mxbai-rerank models",
    "Late-interaction retrieval research",
    "Hosted multimodal search"
   ],
   "focus": [
    "embeddings",
    "retrieval",
    "rerankers"
   ],
   "contributions": [
    {
     "title": "Asymmetric quantization",
     "note": "A June 2026 post describes cutting late-interaction corpus storage 32 times with binary document vectors at near-equal retrieval quality.",
     "year": "2026"
    },
    {
     "title": "Dense retrieval analysis",
     "note": "A June 2026 research post argues dense retrievers encode more information than their scoring mechanism expresses.",
     "year": "2026"
    },
    {
     "title": "Large-file ingestion system",
     "note": "A July 2026 engineering post describes a slicer and parser architecture for files of any size under bounded compute.",
     "year": "2026"
    },
    {
     "title": "mxbai-rerank-v3.1-listwise",
     "note": "A listwise reranker update released in July 2026 inside Mixedbread Search.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "mgrep",
     "url": "https://github.com/mixedbread-ai/mgrep",
     "note": "A calm, CLI-native way to semantically grep everything, like code, images, pdfs and more. (4,439 stars)"
    },
    {
     "name": "baguetter",
     "url": "https://github.com/mixedbread-ai/baguetter",
     "note": "Baguetter is a flexible, efficient, and hackable search engine library implemented in Python. It's designed fo (214 stars)"
    },
    {
     "name": "batched",
     "url": "https://github.com/mixedbread-ai/batched",
     "note": "The Batched API provides a flexible and efficient way to process multiple requests in a batch, with a primary  (161 stars)"
    },
    {
     "name": "nga-art-search",
     "url": "https://github.com/mixedbread-ai/nga-art-search",
     "note": "Semantic Search for over 50,000 images from the National Gallery of Art collection (66 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "mixedbread-ai/mxbai-embed-large-v1",
     "url": "https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1",
     "kind": "model",
     "note": "Model: 2,194,389 downloads, 847 likes."
    },
    {
     "name": "mixedbread-ai/deepset-mxbai-embed-de-large-v1",
     "url": "https://huggingface.co/mixedbread-ai/deepset-mxbai-embed-de-large-v1",
     "kind": "model",
     "note": "Model: 578,018 downloads, 61 likes."
    },
    {
     "name": "mixedbread-ai/baguetter",
     "url": "https://huggingface.co/datasets/mixedbread-ai/baguetter",
     "kind": "dataset",
     "note": "Dataset: 933 downloads, 1 likes."
    },
    {
     "name": "mixedbread-ai/mxbai-rerank-xsmall-v1",
     "url": "https://huggingface.co/mixedbread-ai/mxbai-rerank-xsmall-v1",
     "kind": "model",
     "note": "Model: 541,817 downloads, 59 likes."
    }
   ],
   "papers": [
    {
     "title": "2D Matryoshka Sentence Embeddings",
     "url": "https://arxiv.org/abs/2402.14776",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "mxbai-embed-large-v1",
     "url": "https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1",
     "note": "Open embedding model"
    },
    {
     "name": "mxbai-rerank-large-v2",
     "url": "https://huggingface.co/mixedbread-ai/mxbai-rerank-large-v2",
     "note": "Open reranker"
    },
    {
     "name": "mgrep",
     "url": "https://github.com/mixedbread-ai/mgrep",
     "note": "Semantic grep CLI tool"
    }
   ],
   "openness": "Embedding and reranking models are public on Hugging Face and code is mostly Apache-2.0; the hosted product is closed.",
   "key_people": [
    {
     "name": "Aamir Shakir",
     "role": "Co-founder",
     "url": "https://www.ai.engineer/speakers/aamir-shakir",
     "note": "Co-author of Mixedbread's embedding model papers."
    },
    {
     "name": "Julius Lipp",
     "role": "Co-founder (listed)",
     "url": "https://crunchbase.com/organization/mixedbread-ai",
     "note": "Listed as co-founder by Crunchbase; current role unclear."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website; a support address is printed there.",
   "sources": [
    "https://www.mixedbread.com",
    "https://huggingface.co/mixedbread-ai",
    "https://github.com/mixedbread-ai"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "mixedbread",
   "mlat": 52.54165,
   "mlon": 13.34337,
   "logo": true
  },
  {
   "orig_id": "mideind",
   "name": "Miðeind",
   "alias": null,
   "full_name": "Miðeind",
   "type": "startup",
   "founded": 2015,
   "country": "Iceland",
   "city": "Reykjavik",
   "lat": 64.1466,
   "lon": -21.9426,
   "website": "https://mideind.is",
   "huggingface": "https://huggingface.co/mideind",
   "github": "https://github.com/mideind",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "mideind@mideind.is",
   "summary": "Icelandic language technology company that publishes open Icelandic NLP tools, models and evaluation resources.",
   "tagline": "Reykjavik language-technology company building open-source tools, models and benchmarks so that Icelandic is supported in modern AI.",
   "about": [
    "Miðeind was founded in 2015 by Vilhjálmur Þorsteinsson, a long-time software entrepreneur. It develops language technology and AI for Icelandic, a language whose small market, limited data and complex grammar leave it poorly served by large technology firms.",
    "Its products include Málfríður for spelling and grammar correction, Hreimur for speech recognition, Erlendur for translation, Snara for dictionary search and Lísa for text extraction, all gathered on the Málstaður site. The company states that its software is mostly open source.",
    "Miðeind was an implementation partner in the Icelandic government's Language Technology Programme from 2019 to 2022 and maintains a Hugging Face leaderboard that evaluates large language models on Icelandic."
   ],
   "mission": "Miðeind works to strengthen Icelandic in AI and language technology, doing so as openly as possible so that the public, businesses and researchers all benefit.",
   "notice": null,
   "known_for": [
    "Icelandic spelling, grammar and translation tools",
    "Open-source Icelandic NLP software",
    "Icelandic LLM leaderboard on Hugging Face",
    "Government Language Technology Programme partner"
   ],
   "focus": [
    "Icelandic",
    "NLP",
    "language models",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Vilhjálmur Þorsteinsson founded Miðeind to build language technology for Icelandic.",
     "year": "2015"
    },
    {
     "title": "Language Technology Programme",
     "note": "Miðeind worked as an implementation partner on machine translation, grammar checking and NLP infrastructure under the national programme from 2019 to 2022.",
     "year": "2019"
    },
    {
     "title": "Ský IT Award",
     "note": "The company received the Ský IT Award for outstanding contributions to the Icelandic IT sector.",
     "year": "2024"
    },
    {
     "title": "WMT25 submission",
     "note": "The team published its system description for the WMT25 general machine translation task.",
     "year": "2025"
    },
    {
     "title": "WikiQA-IS benchmark",
     "note": "A benchmark for assisted generation and automated evaluation of Icelandic cultural knowledge in LLMs.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "GreynirEngine",
     "url": "https://github.com/mideind/GreynirEngine",
     "note": "Fast natural language processing engine for Icelandic (70 stars)"
    },
    {
     "name": "GreynirServer",
     "url": "https://github.com/mideind/GreynirServer",
     "note": "The greynir.is Icelandic NLP API and website (67 stars)"
    },
    {
     "name": "Netskrafl",
     "url": "https://github.com/mideind/Netskrafl",
     "note": "Icelandic crossword game website (44 stars)"
    },
    {
     "name": "Tokenizer",
     "url": "https://github.com/mideind/Tokenizer",
     "note": "Tokenizer for Icelandic text (33 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "mideind/IceBERT",
     "url": "https://huggingface.co/mideind/IceBERT",
     "kind": "model",
     "note": "Icelandic RoBERTa model."
    },
    {
     "name": "mideind/yfirlestur-icelandic-correction-byt5",
     "url": "https://huggingface.co/mideind/yfirlestur-icelandic-correction-byt5",
     "kind": "model",
     "note": "Icelandic grammar correction model."
    },
    {
     "name": "mideind/icelandic-error-corpus-IceEC",
     "url": "https://huggingface.co/datasets/mideind/icelandic-error-corpus-IceEC",
     "kind": "dataset",
     "note": "Icelandic error corpus."
    },
    {
     "name": "mideind/icelandic-llm-leaderboard",
     "url": "https://huggingface.co/spaces/mideind/icelandic-llm-leaderboard",
     "kind": "space",
     "note": "Icelandic LLM leaderboard."
    }
   ],
   "papers": [
    {
     "title": "A Warm Start and a Clean Crawled Corpus -- A Recipe for Good Language Models",
     "url": "https://arxiv.org/abs/2201.05601",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "GreynirEngine",
     "url": "https://github.com/mideind/GreynirEngine",
     "note": "Icelandic NLP engine"
    },
    {
     "name": "IceBERT",
     "url": "https://huggingface.co/mideind/IceBERT",
     "note": "Icelandic BERT model"
    },
    {
     "name": "Icelandic LLM Leaderboard",
     "url": "https://huggingface.co/spaces/mideind/icelandic-llm-leaderboard",
     "note": "Leaderboard for Icelandic LLMs"
    },
    {
     "name": "Tokenizer",
     "url": "https://github.com/mideind/Tokenizer",
     "note": "Tokeniser for Icelandic text"
    }
   ],
   "openness": "Tools and models are public on GitHub and Hugging Face with various licences.",
   "key_people": [
    {
     "name": "Vilhjálmur Þorsteinsson",
     "role": "Founder and chairman",
     "url": "https://mideind.is/en/teymid",
     "note": null
    },
    {
     "name": "Linda Heimisdóttir",
     "role": "CEO",
     "url": "https://mideind.is/en/teymid",
     "note": null
    },
    {
     "name": "Þorvaldur Páll Helgason",
     "role": "CTO",
     "url": "https://mideind.is/en/teymid",
     "note": null
    },
    {
     "name": "Hulda Óladóttir",
     "role": "CPO",
     "url": "https://mideind.is/en",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The company site prints general and jobs email addresses.",
   "sources": [
    "https://mideind.is",
    "https://huggingface.co/mideind",
    "https://github.com/mideind"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "mi-eind",
   "mlat": 64.18196,
   "mlon": -21.9426,
   "logo": true
  },
  {
   "orig_id": "mlcommons",
   "name": "MLCommons",
   "alias": null,
   "full_name": "MLCommons",
   "type": "nonprofit",
   "founded": null,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://mlcommons.org",
   "huggingface": "https://huggingface.co/MLCommons",
   "github": "https://github.com/mlcommons",
   "x": "https://x.com/mlcommons",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "MLCommons is a consortium that publishes the MLPerf benchmarks, open datasets such as People's Speech and the AILuminate safety benchmark. Reference implementations are on GitHub and datasets on Hugging Face.",
   "tagline": "A vendor-neutral consortium that builds and publishes the MLPerf benchmarks, open datasets and AI safety measurements used across the industry.",
   "about": [
    "MLCommons grew out of MLPerf, a benchmarking effort begun in 2018 after engineers and researchers from Baidu, Google, Harvard, Stanford and UC Berkeley began meeting to agree on common ways to measure machine learning performance. The MLCommons association was launched in December 2020 to give that work, and related dataset and tooling projects, a permanent home.",
    "Its working groups publish benchmark suites for training, inference, high-performance computing, mobile and tiny devices, together with reference implementations on GitHub. Datasets such as People's Speech and tooling such as MLCube are released openly.",
    "Since 2023 an AI Risk and Reliability group has built safety measurement, which produced the AILuminate benchmark for large language models in December 2024. Member companies, universities and non-profits fund and govern the consortium, and its results are widely cited by hardware vendors and cloud providers."
   ],
   "mission": "MLCommons states its mission as standardizing AI measurement so that buyers, builders and policymakers can make informed decisions about accuracy, safety, speed and efficiency.",
   "notice": null,
   "known_for": [
    "MLPerf training and inference benchmarks",
    "People's Speech open speech dataset",
    "AILuminate LLM safety benchmark",
    "MLCube portable ML packaging"
   ],
   "focus": [
    "evaluation",
    "datasets",
    "benchmarks"
   ],
   "contributions": [
    {
     "title": "MLPerf Training launched",
     "note": "Working group announced the first MLPerf Training benchmark in May after early meetings among industry and academic engineers.",
     "year": "2018"
    },
    {
     "title": "MLPerf Inference launched",
     "note": "The inference benchmark was announced in June, with first results published in November.",
     "year": "2019"
    },
    {
     "title": "MLCube created",
     "note": "A packaging standard for portable machine learning workloads was introduced in November.",
     "year": "2019"
    },
    {
     "title": "MLCommons association launched",
     "note": "The consortium was formally launched on 3 December, absorbing MLPerf and the People's Speech effort.",
     "year": "2020"
    },
    {
     "title": "AI Risk and Reliability group formed",
     "note": "A working group on AI safety measurement began in November.",
     "year": "2023"
    },
    {
     "title": "AILuminate v1.0 released",
     "note": "The first benchmark for assessing the safety of general-purpose chat models was released on 4 December in English.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "training",
     "url": "https://github.com/mlcommons/training",
     "note": "MLPerf training."
    },
    {
     "name": "inference",
     "url": "https://github.com/mlcommons/inference",
     "note": "MLPerf inference."
    },
    {
     "name": "croissant",
     "url": "https://github.com/mlcommons/croissant",
     "note": "Croissant."
    },
    {
     "name": "ck",
     "url": "https://github.com/mlcommons/ck",
     "note": "Collective Knowledge."
    }
   ],
   "hf_highlights": [
    {
     "name": "peoples_speech",
     "url": "https://huggingface.co/datasets/MLCommons/peoples_speech",
     "kind": "dataset",
     "note": "People's Speech."
    },
    {
     "name": "unsupervised_peoples_speech",
     "url": "https://huggingface.co/datasets/MLCommons/unsupervised_peoples_speech",
     "kind": "dataset",
     "note": "Unsup."
    },
    {
     "name": "ml_spoken_words",
     "url": "https://huggingface.co/datasets/MLCommons/ml_spoken_words",
     "kind": "dataset",
     "note": "Spoken Words."
    }
   ],
   "papers": [
    {
     "title": "MLPerf Inference Benchmark",
     "url": "https://arxiv.org/abs/1911.02549",
     "year": "2019"
    },
    {
     "title": "Croissant: A Metadata Format for ML-Ready Datasets",
     "url": "https://arxiv.org/abs/2403.19546",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "MLPerf training",
     "url": "https://github.com/mlcommons/training",
     "note": "Reference benchmarks"
    },
    {
     "name": "MLPerf inference",
     "url": "https://github.com/mlcommons/inference",
     "note": "Reference benchmarks"
    },
    {
     "name": "Croissant",
     "url": "https://github.com/mlcommons/croissant",
     "note": "Dataset metadata format"
    },
    {
     "name": "People's Speech",
     "url": "https://huggingface.co/datasets/MLCommons/peoples_speech",
     "note": "Speech dataset"
    }
   ],
   "openness": "Open code and datasets; licences vary.",
   "key_people": [
    {
     "name": "Peter Mattson",
     "role": "President and founder",
     "url": "https://www.linkedin.com/in/peter-mattson-33b8863/",
     "note": null
    },
    {
     "name": "Rebecca Weiss",
     "role": "Executive Director",
     "url": "https://www.linkedin.com/in/rjweiss/",
     "note": null
    },
    {
     "name": "David Kanter",
     "role": "Head of MLPerf",
     "url": "https://twitter.com/TheKanter",
     "note": null
    },
    {
     "name": "Vijay Janapa Reddi",
     "role": "Vice President",
     "url": "http://scholar.harvard.edu/vijay-janapa-reddi/",
     "note": null
    },
    {
     "name": "Carole-Jean Wu",
     "role": "Vice President",
     "url": "https://sites.google.com/site/carolejeanwu",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://mlcommons.org); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://github.com/mlcommons",
    "https://huggingface.co/MLCommons"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "mlcommons",
   "mlat": 37.87772,
   "mlon": -122.49767,
   "logo": true
  },
  {
   "orig_id": "mlrs-malta",
   "name": "MLRS",
   "alias": "University of Malta",
   "full_name": "MLRS (University of Malta)",
   "type": "academic",
   "founded": null,
   "country": "Malta",
   "city": "Msida",
   "lat": 35.9022,
   "lon": 14.4842,
   "website": "https://mlrs.research.um.edu.mt",
   "huggingface": "https://huggingface.co/MLRS",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Maltese Language Resource Server at the University of Malta, publishing open Maltese language models, corpora and speech data.",
   "tagline": "The University of Malta project that builds open corpora, tools and language models for Maltese, a language with few digital resources.",
   "about": [
    "The Maltese Language Resource Server is a University of Malta project that produces tools and resources for Maltese, English and Maltese Sign Language. It is coordinated by the Institute of Linguistics and Language Technology together with the Department of Artificial Intelligence.",
    "The server hosts Korpus Malti, an annotated corpus of roughly 250 million tokens in its version 3.0, as well as a learner corpus, an open lexicon and a sign language dictionary. A 2016 electronic Maltese dictionary built on the group's lexicon was launched with the Malta Communications Authority.",
    "More recently the group has released Maltese language models and datasets on Hugging Face under the MLRS organisation, for researchers and developers working on Maltese."
   ],
   "mission": "The project aims to supply the open language data and software that Maltese needs to be supported by modern language technology.",
   "notice": null,
   "known_for": [
    "Korpus Malti Maltese text corpus",
    "Open Maltese language models on Hugging Face",
    "Maltese part-of-speech tagging resources",
    "Online Maltese dictionary"
   ],
   "focus": [
    "Maltese NLP"
   ],
   "contributions": [
    {
     "title": "MLRS project described in print",
     "note": "An early paper describes the Maltese Language Resource Server as a two-year research project at the University of Malta.",
     "year": "2006"
    },
    {
     "title": "Korpus Malti v3.0 tagged",
     "note": "Version 3.0 of the corpus was annotated with an SVMTool tagger reporting about 97 percent accuracy.",
     "year": "2016"
    },
    {
     "title": "Official online Maltese dictionary launched",
     "note": "A dictionary built on the group's lexicon was launched with the Malta Communications Authority in March.",
     "year": "2016"
    },
    {
     "title": "Maltese BERT models released",
     "note": "The group published new corpora and BERT-style models for Maltese on Hugging Face.",
     "year": "2022"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "MLRS/BERTu",
     "url": "https://huggingface.co/MLRS/BERTu",
     "kind": "model",
     "note": "Model with 7 likes on the Hub."
    },
    {
     "name": "MLRS/mBERTu",
     "url": "https://huggingface.co/MLRS/mBERTu",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "MLRS/korpus_malti",
     "url": "https://huggingface.co/datasets/MLRS/korpus_malti",
     "kind": "dataset",
     "note": "Dataset with 9 likes on the Hub."
    },
    {
     "name": "MLRS/MELABench",
     "url": "https://huggingface.co/spaces/MLRS/MELABench",
     "kind": "space",
     "note": "Space with 0 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Pre-training Data Quality and Quantity for a Low-Resource Language: New Corpus and BERT Models for Maltese",
     "url": "https://arxiv.org/abs/2205.10517",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "BERTu",
     "url": "https://huggingface.co/MLRS/BERTu",
     "note": "Maltese BERT model"
    },
    {
     "name": "mBERTu",
     "url": "https://huggingface.co/MLRS/mBERTu",
     "note": "Multilingual model adapted to Maltese"
    },
    {
     "name": "Korpus Malti",
     "url": "https://huggingface.co/datasets/MLRS/korpus_malti",
     "note": "Large Maltese text corpus"
    },
    {
     "name": "masri_synthetic",
     "url": "https://huggingface.co/datasets/MLRS/masri_synthetic",
     "note": "Synthetic Maltese speech dataset"
    }
   ],
   "openness": "Models and corpora are published on the MLRS Hugging Face organisation and the MLRS site.",
   "key_people": [
    {
     "name": "Albert Gatt",
     "role": "Collaborator, language technology researcher",
     "url": "https://um.edu.mt/linguistics/ourresearch",
     "note": null
    },
    {
     "name": "Claudia Borg",
     "role": "Researcher, Department of AI",
     "url": "https://um.edu.mt/linguistics/ourresearch",
     "note": null
    },
    {
     "name": "Mike Rosner",
     "role": "Originator of the project",
     "url": "https://live.european-language-grid.eu/ncc/ncc-malta",
     "note": null
    },
    {
     "name": "Kurt Micallef",
     "role": "Researcher on Maltese language models",
     "url": "https://huggingface.co/MLRS",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the site.",
   "sources": [
    "https://mlrs.research.um.edu.mt",
    "https://huggingface.co/MLRS",
    "https://arxiv.org/abs/2205.10517"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "mlrs",
   "logo": true
  },
  {
   "orig_id": "moondream",
   "name": "Moondream",
   "alias": "M87 Labs",
   "full_name": "Moondream (M87 Labs)",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://moondream.ai",
   "huggingface": "https://huggingface.co/moondream",
   "github": "https://github.com/m87-labs",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "contact@moondream.ai",
   "summary": "Company behind Moondream, a compact open vision-language model family for on-device and edge computer vision.",
   "tagline": "Compact open-weight vision-language models for detection, captioning and segmentation, now part of Prime Intellect.",
   "about": [
    "Moondream is a family of small vision-language models built by M87 Labs. The models handle captioning, querying, detection, pointing and segmentation, and run on GPUs, CPUs and edge hardware.",
    "Moondream 3.1 uses a sparse mixture-of-experts design with 9 billion parameters, 2 billion of them active. The site reports more than five million monthly downloads, open weights and commercial use permitted. The company also offers a hosted API, a fine-tuning product and the Photon inference engine.",
    "In October 2026 Moondream announced that it had been acquired by Prime Intellect, which says the combined team will accelerate open models and fast multimodal AI."
   ],
   "mission": "Moondream aims to make capable visual AI small and fast enough to be fine-tuned and deployed on real-world edge cases anywhere.",
   "notice": "Moondream announced on 14 October 2026 that it had been acquired by Prime Intellect, another lab in this atlas.",
   "known_for": [
    "Compact open vision-language models",
    "Moondream 2 and 3 model lines",
    "On-device and edge deployment",
    "Photon inference engine"
   ],
   "focus": [
    "vision-language models",
    "edge inference",
    "computer vision"
   ],
   "contributions": [
    {
     "title": "Moondream 3.1",
     "note": "A mixture-of-experts release with 9 billion parameters and 2 billion active, announced in July 2026.",
     "year": "2026"
    },
    {
     "title": "Photon inference engine",
     "note": "A megakernel inference stack that added speech models and NVIDIA GPU support through 2026.",
     "year": "2026"
    },
    {
     "title": "Parakeet Redux and Ultra",
     "note": "Speech-to-text models for local transcription in 25 languages, released in September 2026.",
     "year": "2026"
    },
    {
     "title": "Acquired by Prime Intellect",
     "note": "Moondream announced on 14 October 2026 that it was joining Prime Intellect.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "moondream",
     "url": "https://github.com/m87-labs/moondream",
     "note": "Tiny vision language model"
    }
   ],
   "hf_highlights": [
    {
     "name": "moondream3-preview",
     "url": "https://huggingface.co/moondream/moondream3-preview",
     "kind": "model",
     "note": "Moondream 3 preview"
    },
    {
     "name": "moondream3.1-9B-A2B",
     "url": "https://huggingface.co/moondream/moondream3.1-9B-A2B",
     "kind": "model",
     "note": "Moondream 3.1 MoE"
    },
    {
     "name": "megalith-mdqa",
     "url": "https://huggingface.co/datasets/moondream/megalith-mdqa",
     "kind": "dataset",
     "note": "Image QA dataset"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "moondream",
     "url": "https://github.com/m87-labs/moondream",
     "note": "Tiny vision language model"
    }
   ],
   "openness": "Apache-2.0 inference code; model weights and datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Vik Korrapati",
     "role": "Founder",
     "url": "https://www.primeintellect.ai/blog/moondream",
     "note": "Joined Prime Intellect to lead inference work after the acquisition."
    },
    {
     "name": "Jay Allen",
     "role": "Founder",
     "url": "https://www.primeintellect.ai/blog/moondream",
     "note": "Joined Prime Intellect alongside Korrapati."
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on website.",
   "sources": [
    "https://moondream.ai/about/us",
    "https://github.com/m87-labs/moondream",
    "https://huggingface.co/moondream"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "moondream",
   "mlat": 37.86202,
   "mlon": -122.5238,
   "logo": true
  },
  {
   "orig_id": "moonshot-ai",
   "name": "Moonshot AI",
   "alias": "Kimi",
   "full_name": "Moonshot AI (Kimi)",
   "type": "startup",
   "founded": 2023,
   "country": "China",
   "city": "Beijing",
   "lat": 39.9,
   "lon": 116.41,
   "website": "https://www.moonshot.ai",
   "huggingface": "https://huggingface.co/moonshotai",
   "github": "https://github.com/moonshotai",
   "x": "https://twitter.com/Kimi_Moonshot",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Beijing company behind the Kimi assistant and model family. It publishes Kimi model weights on Hugging Face.",
   "tagline": "Beijing lab behind the Kimi models, whose open-weight Kimi K3 family at 2.8 trillion parameters is the largest open release to date.",
   "about": [
    "Moonshot AI was founded in Beijing in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin, who had been schoolfriends at Tsinghua University. Yang previously worked at Google Brain and Meta. The company is named after Pink Floyd's The Dark Side of the Moon.",
    "It operates the Kimi assistant, an API platform and the Kimi Code command-line agent, and publishes model weights on Hugging Face. Kimi K2 appeared in July 2025, K2.5 added native vision in January 2026, K3 followed in mid-2026 and Kimi K3.1 Pro was published in November 2026.",
    "Moonshot is venture-backed, with Alibaba and Tencent among its investors, and releases K3 under a custom license that requires revenue sharing from very large commercial inference providers."
   ],
   "mission": "Moonshot states its aim as building foundation models toward artificial general intelligence, sharing research with the open-source community.",
   "notice": null,
   "known_for": [
    "Kimi open-weight model family",
    "Very large mixture-of-experts open models",
    "Long-context chat assistant",
    "Agent Swarm and Kimi Code"
   ],
   "focus": [
    "language models",
    "multimodal",
    "agents",
    "Chinese",
    "English"
   ],
   "contributions": [
    {
     "title": "Kimi chatbot",
     "note": "Moonshot released the first version of Kimi in October 2023, capable of long Chinese-language context.",
     "year": "2023"
    },
    {
     "title": "Kimi K2",
     "note": "The K2 weights were published on Hugging Face in July 2025.",
     "year": "2025"
    },
    {
     "title": "Kimi K2.5",
     "note": "A multimodal upgrade with a 400-million-parameter MoonViT vision encoder and Agent Swarm, released in January 2026.",
     "year": "2026"
    },
    {
     "title": "Kimi K3",
     "note": "A 2.8-trillion-parameter open-weight model family launched in July 2026.",
     "year": "2026"
    },
    {
     "title": "Kimi K3.1 Pro",
     "note": "A model with a one-million-token context window published in October and November 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "kimi-cli",
     "url": "https://github.com/MoonshotAI/kimi-cli",
     "note": "Archived legacy Kimi CLI, the most starred repo."
    },
    {
     "name": "Kimi-K2",
     "url": "https://github.com/MoonshotAI/Kimi-K2",
     "note": "Repository for the Kimi K2 model series."
    },
    {
     "name": "Kimi-K3",
     "url": "https://github.com/MoonshotAI/Kimi-K3",
     "note": "Repository for Kimi K3."
    },
    {
     "name": "kimi-code",
     "url": "https://github.com/MoonshotAI/kimi-code",
     "note": "Kimi Code CLI for agentic coding."
    }
   ],
   "hf_highlights": [
    {
     "name": "Kimi-K3",
     "url": "https://huggingface.co/moonshotai/Kimi-K3",
     "kind": "model",
     "note": "Most downloaded model in the org."
    },
    {
     "name": "Kimi-K2.5",
     "url": "https://huggingface.co/moonshotai/Kimi-K2.5",
     "kind": "model",
     "note": "Kimi K2.5 release."
    },
    {
     "name": "Kimi-K2-Instruct",
     "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct",
     "kind": "model",
     "note": "Instruction-tuned K2."
    },
    {
     "name": "PerceptionBench",
     "url": "https://huggingface.co/datasets/moonshotai/PerceptionBench",
     "kind": "dataset",
     "note": "Multimodal perception benchmark."
    }
   ],
   "papers": [
    {
     "title": "Kimi K2: Open Agentic Intelligence",
     "url": "https://arxiv.org/abs/2507.20534",
     "year": "2025"
    },
    {
     "title": "Kimi k1.5: Scaling Reinforcement Learning with LLMs",
     "url": "https://arxiv.org/abs/2501.12599",
     "year": "2025"
    },
    {
     "title": "Muon is Scalable for LLM Training",
     "url": "https://arxiv.org/abs/2502.16982",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Kimi-K3",
     "url": "https://huggingface.co/moonshotai/Kimi-K3",
     "note": "Kimi K3 model"
    },
    {
     "name": "Kimi-K3.1-Pro",
     "url": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro",
     "note": "Kimi K3.1 Pro"
    },
    {
     "name": "Kimi-K2.6",
     "url": "https://huggingface.co/moonshotai/Kimi-K2.6",
     "note": "Kimi K2.6"
    }
   ],
   "openness": "Open weights on Hugging Face; exact licence stated on each model card.",
   "key_people": [
    {
     "name": "Yang Zhilin",
     "role": "Co-founder and CEO",
     "url": "https://en.wikipedia.org/wiki/Moonshot_AI",
     "note": null
    },
    {
     "name": "Zhou Xinyu",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Moonshot_AI",
     "note": null
    },
    {
     "name": "Wu Yuxin",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Moonshot_AI",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Reach the team through its website, GitHub issues or Hugging Face discussions.",
   "sources": [
    "https://huggingface.co/moonshotai",
    "https://www.moonshot.ai"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "moonshot-ai",
   "mlat": 39.88273,
   "mlon": 116.4793,
   "logo": true
  },
  {
   "orig_id": "motif-technologies",
   "name": "Motif Technologies",
   "alias": null,
   "full_name": "Motif Technologies",
   "type": "startup",
   "founded": 2025,
   "country": "South Korea",
   "city": "Seoul",
   "lat": 37.57,
   "lon": 126.98,
   "website": "https://motiftech.io/",
   "huggingface": "https://huggingface.co/Motif-Technologies",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Korean AI company publishing Motif base models and open optimizer and activation kernels.",
   "tagline": "Korean AI lab, a spin-off of the infrastructure firm Moreh, publishing open-weight Motif language models and the training techniques behind them.",
   "about": [
    "Motif Technologies is a Korean AI company reported to have been established in February 2025 as a spin-off from the AI infrastructure company Moreh. It trains its own foundation models, led by CEO Junghwan Lim according to press reports and its technical report.",
    "Its Hugging Face organization carries the Motif model family, including the Motif-2-12.7B open-weight model described in a November 2025 technical report, as well as audio, vision encoder, video and VAE components. The report describes a new attention design, a curriculum-based data schedule and optimized training kernels.",
    "Motif-3, a sparse mixture-of-experts model with about 314 billion parameters, was published in August 2026 under a non-commercial research license, a departure from the Apache 2.0 license used for Motif-2. Press reports link its training to South Korea's national sovereign-AI program."
   ],
   "mission": "Motif aims to build competitive, domestically developed foundation models and the efficient training methods behind them.",
   "notice": null,
   "known_for": [
    "Motif-2 and Motif-3 open-weight models",
    "Efficiency-focused training kernels and optimizers",
    "Korean sovereign-AI program participant"
   ],
   "focus": [
    "language models",
    "training efficiency",
    "Korean"
   ],
   "contributions": [
    {
     "title": "Motif 2.6B",
     "note": "A 2.6-billion-parameter small language model published on Hugging Face in 2025.",
     "year": "2025"
    },
    {
     "title": "Motif-2-12.7B technical report",
     "note": "The November 2025 report presents an open-weight model with a new attention design, curriculum data schedule and optimized kernels.",
     "year": "2025"
    },
    {
     "title": "Motif-3 preview",
     "note": "A roughly 314-billion-parameter mixture-of-experts beta checkpoint appeared on Hugging Face in July 2026.",
     "year": "2026"
    },
    {
     "title": "Motif-3 release",
     "note": "The Motif-3 and Motif-3-Base weights were published in August 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "Motif-Technologies/Motif-3",
     "url": "https://huggingface.co/Motif-Technologies/Motif-3",
     "kind": "model",
     "note": "Most downloaded Motif model."
    },
    {
     "name": "Motif-Technologies/Motif-Video-2B-GGUF",
     "url": "https://huggingface.co/Motif-Technologies/Motif-Video-2B-GGUF",
     "kind": "model",
     "note": "GGUF build of the video model."
    },
    {
     "name": "Motif-Technologies/Motif-2.6B",
     "url": "https://huggingface.co/Motif-Technologies/Motif-2.6B",
     "kind": "model",
     "note": "2.6B base model."
    },
    {
     "name": "Motif-Technologies/agenttracer",
     "url": "https://huggingface.co/datasets/Motif-Technologies/agenttracer",
     "kind": "dataset",
     "note": "Agent trace dataset."
    }
   ],
   "papers": [
    {
     "title": "Motif 2.6B Technical Report",
     "url": "https://arxiv.org/abs/2508.09148",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Motif-3-Base",
     "url": "https://huggingface.co/Motif-Technologies/Motif-3-Base",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "optimizer",
     "url": "https://huggingface.co/Motif-Technologies/optimizer",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights and kernels on Hugging Face.",
   "key_people": [
    {
     "name": "Junghwan Lim",
     "role": "CEO",
     "url": "https://arxiv.org/abs/2511.07464",
     "note": "Lead author of the Motif-2-12.7B technical report; previously AI director at Moreh."
    }
   ],
   "discord_note": null,
   "contact_note": "The website motiftech.io is minimal; Hugging Face is the main public channel.",
   "sources": [
    "https://huggingface.co/Motif-Technologies",
    "https://motiftech.io/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "motif-technologies",
   "mlat": 37.5267,
   "mlon": 126.98,
   "logo": true
  },
  {
   "orig_id": "mozilla-ai",
   "name": "Mozilla.ai",
   "alias": null,
   "full_name": "Mozilla.ai",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.mozilla.ai",
   "huggingface": "https://huggingface.co/mozilla-ai",
   "github": "https://github.com/mozilla-ai",
   "x": null,
   "discord": "https://discord.gg/4gf3zXrQUc",
   "linkedin": null,
   "contact_email": null,
   "summary": "Public benefit company founded by Mozilla (a subsidiary of the Mozilla Foundation) building open-source tools for trustworthy, user-controlled AI such as llamafile, any-llm and any-agent.",
   "tagline": "Mozilla-backed company building open-source tools that give people and organizations choice, transparency and control over how AI is used.",
   "about": [
    "Mozilla.ai is a distributed company backed by the Mozilla Foundation that builds open-source AI products and infrastructure. Its GitHub organization dates from April 2023.",
    "Its tools include any-llm, a single interface to many model providers, any-agent for running and evaluating agent frameworks, any-guardrail, mcpd, encoderfile and the Lumigator evaluation platform. It also hosts llamafile, the project that packages a model as one executable. Its newer products are Otari, an open-source governance layer for AI agents, and Octonous.",
    "The team values open foundations and keeps humans in charge of outcomes; the code is largely released under Apache 2.0."
   ],
   "mission": "Mozilla.ai builds open-source AI products and infrastructure that give people and organizations more choice, transparency and control.",
   "notice": "Mozilla.ai is a subsidiary of the Mozilla Foundation; its code is open source.",
   "known_for": [
    "llamafile single-file local models",
    "any-llm and any-agent interfaces",
    "Lumigator model evaluation",
    "Otari agent governance"
   ],
   "focus": [
    "tooling",
    "agents",
    "evaluation",
    "local inference"
   ],
   "contributions": [
    {
     "title": "Mozilla.ai GitHub organization",
     "note": "The public organization was created in April 2023.",
     "year": "2023"
    },
    {
     "title": "llamafile repository",
     "note": "The llamafile project, which bundles a model and runtime into one executable, was created in September 2023 and is now hosted under mozilla-ai.",
     "year": "2023"
    },
    {
     "title": "Lumigator",
     "note": "A platform for evaluating language models, first published in March 2024.",
     "year": "2024"
    },
    {
     "title": "any-agent",
     "note": "A single interface for using and evaluating different agent frameworks, created in March 2025.",
     "year": "2025"
    },
    {
     "title": "any-llm",
     "note": "A unified interface to LLM providers, created in July 2025.",
     "year": "2025"
    },
    {
     "title": "encoderfile",
     "note": "A tool to distribute and run transformer encoders as a single file, created in October 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "llamafile",
     "url": "https://github.com/mozilla-ai/llamafile",
     "note": "Single-file LLMs."
    },
    {
     "name": "any-llm",
     "url": "https://github.com/mozilla-ai/any-llm",
     "note": "Single interface for LLM providers."
    },
    {
     "name": "cq",
     "url": "https://github.com/mozilla-ai/cq",
     "note": "Shared agent learning standard."
    },
    {
     "name": "any-agent",
     "url": "https://github.com/mozilla-ai/any-agent",
     "note": "Agent framework interface."
    }
   ],
   "hf_highlights": [
    {
     "name": "Llama-3.2-1B-Instruct-llamafile",
     "url": "https://huggingface.co/mozilla-ai/Llama-3.2-1B-Instruct-llamafile",
     "kind": "model",
     "note": "Llamafile."
    },
    {
     "name": "Meta-Llama-3-8B-Instruct-llamafile",
     "url": "https://huggingface.co/mozilla-ai/Meta-Llama-3-8B-Instruct-llamafile",
     "kind": "model",
     "note": "Llamafile 8B."
    },
    {
     "name": "TinyLlama-1.1B-Chat-v1.0-llamafile",
     "url": "https://huggingface.co/mozilla-ai/TinyLlama-1.1B-Chat-v1.0-llamafile",
     "kind": "model",
     "note": "TinyLlama llamafile."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "llamafile",
     "url": "https://github.com/mozilla-ai/llamafile",
     "note": "Single-file LLM distribution"
    },
    {
     "name": "any-llm",
     "url": "https://github.com/mozilla-ai/any-llm",
     "note": "Unified LLM provider interface"
    },
    {
     "name": "any-agent",
     "url": "https://github.com/mozilla-ai/any-agent",
     "note": "Agent framework interface"
    }
   ],
   "openness": "Open-source code, mostly permissive licences; affiliated with the Mozilla Foundation.",
   "key_people": [
    {
     "name": "Alejandro Gonzalez",
     "role": "VP, Engineering",
     "url": "https://github.com/agpituk",
     "note": null
    },
    {
     "name": "Davide Eynard",
     "role": "Engineer, public org member",
     "url": "http://davide.eynard.it",
     "note": null
    },
    {
     "name": "Sally Coome",
     "role": "Chief of Staff",
     "url": "https://www.mozilla.ai/company/about-us",
     "note": null
    },
    {
     "name": "Carla Smith",
     "role": "Head of Finance",
     "url": "https://www.mozilla.ai/company/about-us",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.mozilla.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.mozilla.ai",
    "https://huggingface.co/mozilla-ai",
    "https://github.com/mozilla-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "mozilla-ai",
   "mlat": 37.84224,
   "mlon": -122.54506,
   "logo": true
  },
  {
   "orig_id": "mtg-upf",
   "name": "Music Technology Group",
   "alias": "UPF",
   "full_name": "Music Technology Group (UPF)",
   "type": "academic",
   "founded": 1994,
   "country": "Spain",
   "city": "Barcelona",
   "lat": 41.3874,
   "lon": 2.1686,
   "website": "https://www.upf.edu/web/mtg",
   "huggingface": null,
   "github": "https://github.com/MTG",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Music Technology Group at Universitat Pompeu Fabra, a research group building open-source audio and music analysis software including the Essentia library.",
   "tagline": "The Barcelona research group behind Essentia and Freesound-linked work, building open audio and music analysis software since 1994.",
   "about": [
    "The Music Technology Group is a research group of the Department of Information and Communication Technologies at Universitat Pompeu Fabra in Barcelona. It was founded in 1994 by Xavier Serra, who still directs it.",
    "Its research covers audio signal processing, sound and music description, musical interaction, and sound and music communities. The group releases much of its software as open source on GitHub.",
    "Its best known software is Essentia, a C++ library with Python bindings for audio analysis and music information retrieval. The library began in 2006 to replace a set of ad hoc C++ programs and is released under the AGPLv3 licence."
   ],
   "mission": "The group works to advance sound and music computing and to turn that research into open tools used by scientists, developers and artists.",
   "notice": null,
   "known_for": [
    "Essentia audio analysis library",
    "Founded 1994 by Xavier Serra",
    "Sound and music computing research",
    "Open music information retrieval datasets and tools"
   ],
   "focus": [
    "music AI",
    "audio analysis"
   ],
   "contributions": [
    {
     "title": "MTG founded",
     "note": "Xavier Serra founded the group at Universitat Pompeu Fabra.",
     "year": "1994"
    },
    {
     "title": "Essentia development begins",
     "note": "Started to replace ad hoc C++ programs used inside the group.",
     "year": "2006"
    },
    {
     "title": "Essentia 1.0",
     "note": "First numbered release of the library, per Essentia documentation.",
     "year": "2008"
    },
    {
     "title": "Essentia paper at ACM Multimedia",
     "note": "Bogdanov and colleagues presented the open-source library at MM'13.",
     "year": "2013"
    }
   ],
   "github_highlights": [
    {
     "name": "essentia",
     "url": "https://github.com/MTG/essentia",
     "note": "C++ library for audio and music analysis, description and synthesis, including Python bindings (3,808 stars)"
    },
    {
     "name": "sms-tools",
     "url": "https://github.com/MTG/sms-tools",
     "note": "Sound analysis/synthesis tools for music applications (1,801 stars)"
    },
    {
     "name": "essentia.js",
     "url": "https://github.com/MTG/essentia.js",
     "note": "JavaScript library for music/audio analysis and processing powered by Essentia WebAssembly (874 stars)"
    },
    {
     "name": "DeepConvSep",
     "url": "https://github.com/MTG/DeepConvSep",
     "note": "Deep Convolutional Neural Networks for Musical Source Separation  (484 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Essentia",
     "url": "https://github.com/MTG/essentia",
     "note": "C++ library for audio and music analysis with Python bindings"
    },
    {
     "name": "sms-tools",
     "url": "https://github.com/MTG/sms-tools",
     "note": "Sound analysis and synthesis tools for music applications"
    },
    {
     "name": "essentia.js",
     "url": "https://github.com/MTG/essentia.js",
     "note": "WebAssembly port of Essentia for the browser"
    },
    {
     "name": "DeepConvSep",
     "url": "https://github.com/MTG/DeepConvSep",
     "note": "Deep convolutional networks for musical source separation"
    }
   ],
   "openness": "Libraries and research code are public on GitHub; Essentia is released as open source (AGPL with commercial licences available).",
   "key_people": [
    {
     "name": "Xavier Serra",
     "role": "Founder and director",
     "url": "https://en.wikipedia.org/wiki/Music_Technology_Group",
     "note": null
    },
    {
     "name": "Emilia Gomez",
     "role": "Academic staff",
     "url": "https://en.wikipedia.org/wiki/Music_Technology_Group",
     "note": null
    },
    {
     "name": "Sergi Jorda",
     "role": "Academic staff",
     "url": "https://en.wikipedia.org/wiki/Music_Technology_Group",
     "note": null
    },
    {
     "name": "Rafael Ramirez",
     "role": "Academic staff",
     "url": "https://en.wikipedia.org/wiki/Music_Technology_Group",
     "note": null
    },
    {
     "name": "Dmitry Bogdanov",
     "role": "Essentia lead author",
     "url": "https://essentia.upf.edu/introduction.html",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website.",
   "sources": [
    "https://www.upf.edu/web/mtg",
    "https://github.com/MTG"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "music-technology-group",
   "mlat": 41.3874,
   "mlon": 2.23524,
   "logo": true
  },
  {
   "orig_id": "nbailab",
   "name": "NB AI-Lab",
   "alias": "National Library of Norway",
   "full_name": "NB AI-Lab (National Library of Norway)",
   "type": "institute",
   "founded": null,
   "country": "Norway",
   "city": "Mo i Rana",
   "lat": 66.3128,
   "lon": 14.1428,
   "website": "https://ai.nb.no",
   "huggingface": "https://huggingface.co/NbAiLab",
   "github": "https://github.com/NbAiLab",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "ai-lab@nb.no",
   "summary": "AI lab at the National Library of Norway that trains and releases Norwegian language and speech models from the library's digital collections.",
   "tagline": "The National Library of Norway's AI lab, which releases Norwegian language models, speech models and datasets built from the library's collections.",
   "about": [
    "The AI-lab at the National Library of Norway explores how AI can serve libraries, archives and museums, and builds datasets and models for the wider community. It is a unit of the library rather than a separate company.",
    "Its first major output, the Norwegian Transformer Model Project of 2020 to 2021, produced nb-bert-base, described as the first working Norwegian transformer model, and the Norwegian Colossal Corpus. A later Norwegian Speech Transformer Project, led by Per Egil Kummervold from September 2022, targeted better Norwegian speech-to-text.",
    "The lab publishes its work through its GitHub and Hugging Face organizations."
   ],
   "mission": "The lab pursues applications of AI in libraries, archives and museums and contributes datasets and models so that the library can help shape good AI for Norwegian.",
   "notice": null,
   "known_for": [
    "nb-bert-base, the first Norwegian transformer model",
    "Norwegian Colossal Corpus",
    "Norwegian speech models"
   ],
   "focus": [
    "Norwegian",
    "speech",
    "language models",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Norwegian Transformer Model Project",
     "note": "A project running from September 2020 to December 2021 that produced Norwegian BERT models and the Norwegian Colossal Corpus.",
     "year": "2020"
    },
    {
     "title": "nb-bert-base",
     "note": "The first working Norwegian transformer model, published during the Norwegian Transformer Model Project.",
     "year": "2021"
    },
    {
     "title": "Norwegian Speech Transformer Project",
     "note": "A project led by Per Egil Kummervold from September 2022 to improve Norwegian speech-to-text with transformers.",
     "year": "2022"
    }
   ],
   "github_highlights": [
    {
     "name": "notram",
     "url": "https://github.com/NbAiLab/notram",
     "note": "Norwegian Transformer Model (123 stars)"
    },
    {
     "name": "nb-whisper",
     "url": "https://github.com/NbAiLab/nb-whisper",
     "note": "Open-source repository (58 stars)"
    },
    {
     "name": "nostram",
     "url": "https://github.com/NbAiLab/nostram",
     "note": "Norwegian Speech Transformer Models (19 stars)"
    },
    {
     "name": "tokenizer-benchmark",
     "url": "https://github.com/NbAiLab/tokenizer-benchmark",
     "note": "Benchmark for Scandinavian Tokenizers (9 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "NbAiLab/nb-wav2vec2-1b-nynorsk",
     "url": "https://huggingface.co/NbAiLab/nb-wav2vec2-1b-nynorsk",
     "kind": "model",
     "note": "Model: 500,530 downloads, 4 likes."
    },
    {
     "name": "NbAiLab/nb-wav2vec2-1b-bokmaal-v2",
     "url": "https://huggingface.co/NbAiLab/nb-wav2vec2-1b-bokmaal-v2",
     "kind": "model",
     "note": "Model: 444,207 downloads, 0 likes."
    },
    {
     "name": "NbAiLab/NCC",
     "url": "https://huggingface.co/datasets/NbAiLab/NCC",
     "kind": "dataset",
     "note": "Dataset: 1,617 downloads, 4 likes."
    },
    {
     "name": "NbAiLab/nb-tts-voxcpm2-voices-2607-demo",
     "url": "https://huggingface.co/spaces/NbAiLab/nb-tts-voxcpm2-voices-2607-demo",
     "kind": "space",
     "note": "Space: 8 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Operationalizing a National Digital Library: The Case for a Norwegian Transformer Model",
     "url": "https://arxiv.org/abs/2104.09617",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "NB-Whisper large",
     "url": "https://huggingface.co/NbAiLab/nb-whisper-large",
     "note": "Norwegian speech recognition model"
    },
    {
     "name": "NB-BERT base",
     "url": "https://huggingface.co/NbAiLab/nb-bert-base",
     "note": "Norwegian BERT model"
    },
    {
     "name": "NPSC",
     "url": "https://huggingface.co/datasets/NbAiLab/NPSC",
     "note": "Norwegian Parliamentary Speech Corpus"
    },
    {
     "name": "notram",
     "url": "https://github.com/NbAiLab/notram",
     "note": "Norwegian transformer models"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face; code is mainly Apache-2.0.",
   "key_people": [
    {
     "name": "Javier de la Rosa",
     "role": "Leads AI activity; Research Scientist",
     "url": "http://versae.es",
     "note": null
    },
    {
     "name": "Per Egil Kummervold",
     "role": "Senior Researcher",
     "url": "https://ai.nb.no/about-us/",
     "note": null
    },
    {
     "name": "Freddy Wetjen",
     "role": "Responsible for early AI development at the library",
     "url": "https://ai.nb.no/about-us/",
     "note": null
    },
    {
     "name": "Rolv-Arild Braaten",
     "role": "Builds tools for internal and public use",
     "url": "https://ai.nb.no/about-us/",
     "note": null
    },
    {
     "name": "Svein Arne Brygfjeld",
     "role": "Advisor on AI",
     "url": "https://ai.nb.no/about-us/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on the website.",
   "sources": [
    "https://ai.nb.no",
    "https://huggingface.co/NbAiLab",
    "https://github.com/NbAiLab"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "nb-ai-lab",
   "logo": true
  },
  {
   "orig_id": "neuphonic",
   "name": "Neuphonic",
   "alias": null,
   "full_name": "Neuphonic",
   "type": "startup",
   "founded": null,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://www.neuphonic.com",
   "huggingface": "https://huggingface.co/neuphonic",
   "github": "https://github.com/neuphonic",
   "x": "https://x.com/neuphonicspeech",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/neuphonic",
   "contact_email": "general@neuphonic.com",
   "summary": "UK voice-AI lab building text-to-speech models. It releases on-device TTS models and audio codecs openly.",
   "tagline": "London voice-AI company building low-latency text-to-speech and publishing open on-device NeuTTS models and the NeuCodec audio codec.",
   "about": [
    "Neuphonic is a text-to-speech company based in London. Its co-founders, who met studying machine learning at Cambridge, are Sohaib Ahmad as CEO and Jiameng Gao as CTO, according to press coverage.",
    "It builds ultra-low-latency speech that can run locally, on premises or through an API, with multilingual voices and instant voice cloning. Its site lists the NeuTTS family and the NeuCodec audio codec and links to its GitHub and Hugging Face organizations.",
    "The company has reported more than one million open-source downloads and says it is partnering with device makers."
   ],
   "mission": "Neuphonic aims to make AI voice interaction immediate and natural through fast, private text-to-speech that developers can run anywhere.",
   "notice": null,
   "known_for": [
    "NeuTTS on-device speech models",
    "NeuCodec audio codec",
    "Low-latency text-to-speech"
   ],
   "focus": [
    "speech",
    "text-to-speech",
    "audio codecs"
   ],
   "contributions": [
    {
     "title": "Neuphonic GitHub organization",
     "note": "The public GitHub organization was created in April 2024.",
     "year": "2024"
    },
    {
     "title": "NeuTTS family",
     "note": "Open on-device speech language models with instant voice cloning, published under Apache 2.0 or the NeuTTS Open License depending on the model.",
     "year": "2025"
    },
    {
     "title": "One million open-source downloads",
     "note": "The CEO cited more than one million open-source downloads in an April 2026 press release.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "neutts",
     "url": "https://github.com/neuphonic/neutts",
     "note": "On-device TTS model (6,340 stars)"
    },
    {
     "name": "neucodec",
     "url": "https://github.com/neuphonic/neucodec",
     "note": "50 Hz, 0.8 kbps, 24 kHz audio codec (170 stars)"
    },
    {
     "name": "neudecide",
     "url": "https://github.com/neuphonic/neudecide",
     "note": "Tiny voice-action model for any device (102 stars)"
    },
    {
     "name": "pyneuphonic",
     "url": "https://github.com/neuphonic/pyneuphonic",
     "note": "Python SDK for the Neuphonic TTS engine (12 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "neuphonic/neucodec",
     "url": "https://huggingface.co/neuphonic/neucodec",
     "kind": "model",
     "note": "Audio codec, the group's most downloaded model."
    },
    {
     "name": "neuphonic/neutts-air",
     "url": "https://huggingface.co/neuphonic/neutts-air",
     "kind": "model",
     "note": "NeuTTS Air weights."
    },
    {
     "name": "neuphonic/neutts-air-q4-gguf",
     "url": "https://huggingface.co/neuphonic/neutts-air-q4-gguf",
     "kind": "model",
     "note": "Quantised GGUF NeuTTS Air."
    },
    {
     "name": "neuphonic/neutts-nano-multilingual-collection",
     "url": "https://huggingface.co/spaces/neuphonic/neutts-nano-multilingual-collection",
     "kind": "space",
     "note": "Multilingual NeuTTS Nano collection demo."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "NeuTTS",
     "url": "https://github.com/neuphonic/neutts",
     "note": "On-device TTS model"
    },
    {
     "name": "neutts-air",
     "url": "https://huggingface.co/neuphonic/neutts-air",
     "note": "Open TTS model"
    },
    {
     "name": "NeuCodec",
     "url": "https://huggingface.co/neuphonic/neucodec",
     "note": "Audio codec"
    }
   ],
   "openness": "Code and models are public on GitHub and Hugging Face; the NeuTTS licence is not standard, so check the repository.",
   "key_people": [
    {
     "name": "Sohaib Ahmad",
     "role": "Co-founder and CEO",
     "url": "https://www.neuphonic.com",
     "note": null
    },
    {
     "name": "Jiameng Gao",
     "role": "Co-founder and CTO",
     "url": "https://www.neuphonic.com",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Its contact page lists a general email address.",
   "sources": [
    "https://www.neuphonic.com",
    "https://huggingface.co/neuphonic",
    "https://github.com/neuphonic"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "neuphonic",
   "mlat": 51.56699,
   "mlon": -0.08169,
   "logo": true
  },
  {
   "orig_id": "neuronpedia",
   "name": "Neuronpedia",
   "alias": null,
   "full_name": "Neuronpedia",
   "type": "nonprofit",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.neuronpedia.org",
   "huggingface": "https://huggingface.co/neuronpedia",
   "github": "https://github.com/hijohnnylin",
   "x": "https://x.com/neuronpedia",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source interpretability platform for exploring, visualising and steering model internals such as sparse autoencoder features.",
   "tagline": "An open-source interpretability platform for exploring, visualizing and steering model internals such as sparse autoencoder features.",
   "about": [
    "Neuronpedia was created by Johnny Lin, an engineer who previously worked at Apple. It is an open-source platform where researchers and the public can search the features inside language models, steer model behavior and trace circuits.",
    "It hosts feature dashboards and tooling for releases such as Gemma Scope 2, Llama Scope and SAE Bench, and recent additions include Circuit Tracer, the Assistant Axis, Natural Language Autoencoders and Jacobian Lens. It also offers an API and a community Slack.",
    "Funding comes from Decode Research, Open Philanthropy, the Long Term Future Fund and others, and its users include groups at Anthropic, Google DeepMind, EleutherAI and Apollo Research."
   ],
   "mission": "Neuronpedia aims to make model internals easy to inspect and share, so that interpretability research is open and collaborative.",
   "notice": null,
   "known_for": [
    "Public sparse autoencoder feature explorer",
    "Hosting Gemma Scope and Llama Scope",
    "Circuit Tracer and model steering tools",
    "Open-source interp-engine"
   ],
   "focus": [
    "interpretability",
    "sparse autoencoders",
    "tooling"
   ],
   "contributions": [
    {
     "title": "Neuronpedia created",
     "note": "Started as a game for labeling neurons.",
     "year": "2023"
    },
    {
     "title": "Refocus on researchers",
     "note": "Shifted to supporting interpretability researchers.",
     "year": "2024"
    },
    {
     "title": "Open-sourced",
     "note": "Released the platform's code publicly on 31 March 2025.",
     "year": "2025"
    },
    {
     "title": "Circuit Tracer",
     "note": "Hosted the open circuit-tracing integration with Anthropic and partners.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "neuronpedia",
     "url": "https://github.com/hijohnnylin/neuronpedia",
     "note": "Open source interpretability platform"
    },
    {
     "name": "neuronpedia-python",
     "url": "https://github.com/hijohnnylin/neuronpedia-python",
     "note": "Python client"
    }
   ],
   "hf_highlights": [
    {
     "name": "dashboards",
     "url": "https://huggingface.co/datasets/neuronpedia/dashboards",
     "kind": "dataset",
     "note": "Feature dashboards"
    },
    {
     "name": "jacobian-lens",
     "url": "https://huggingface.co/neuronpedia/jacobian-lens",
     "kind": "model",
     "note": "Interpretability artefact"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "neuronpedia",
     "url": "https://github.com/hijohnnylin/neuronpedia",
     "note": "Open source interpretability platform"
    },
    {
     "name": "neuronpedia-python",
     "url": "https://github.com/hijohnnylin/neuronpedia-python",
     "note": "Python client"
    }
   ],
   "openness": "Apache-2.0 platform code with open dashboards data on Hugging Face.",
   "key_people": [
    {
     "name": "Johnny Lin",
     "role": "Founder and lead",
     "url": "https://johnnylin.co",
     "note": "Creator of Neuronpedia."
    },
    {
     "name": "Curt Tigges",
     "role": "Science lead, Decode Research",
     "url": "https://www.anthropic.com/research/open-source-circuit-tracing",
     "note": "Named by Anthropic as science lead of the Neuronpedia integration."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.neuronpedia.org/",
    "https://github.com/hijohnnylin/neuronpedia",
    "https://huggingface.co/neuronpedia"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "neuronpedia",
   "mlat": 37.81932,
   "mlon": -122.56043,
   "logo": true
  },
  {
   "orig_id": "nlx-ulisboa",
   "name": "NLX Group",
   "alias": "University of Lisbon, PORTULAN",
   "full_name": "NLX Group (University of Lisbon, PORTULAN)",
   "type": "academic",
   "founded": null,
   "country": "Portugal",
   "city": "Lisbon",
   "lat": 38.7223,
   "lon": -9.1393,
   "website": "http://nlx.di.fc.ul.pt",
   "huggingface": "https://huggingface.co/PORTULAN",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Natural Language and Speech group at the University of Lisbon, creators of Albertina and Gervasio Portuguese language models.",
   "tagline": "The University of Lisbon language technology group that released the Albertina and Gervasio open models for Portuguese.",
   "about": [
    "NLX is the Natural Language and Speech Group at the Department of Informatics of the Faculty of Sciences, University of Lisbon, led by António Branco. Together with the PORTULAN CLARIN infrastructure it publishes language resources and tools for Portuguese.",
    "Its Hugging Face organisation PORTULAN hosts the Albertina encoder family in European and Brazilian Portuguese, from 100 million to 1.5 billion parameters, and the Gervasio decoder models at 8 and 70 billion parameters.",
    "The group also publishes benchmark datasets such as extraglue and glue-ptpt and maintains a leaderboard of open language models for Portuguese."
   ],
   "mission": "The group works to give Portuguese, in both its European and Brazilian varieties, open neural models and evaluation resources of high quality.",
   "notice": null,
   "known_for": [
    "Albertina Portuguese encoder models",
    "Gervasio open Portuguese decoders",
    "PORTULAN CLARIN infrastructure",
    "Portuguese GLUE-style benchmarks"
   ],
   "focus": [
    "Portuguese NLP",
    "language models"
   ],
   "contributions": [
    {
     "title": "Albertina PT-* introduced",
     "note": "DeBERTa-based encoders for European and Brazilian Portuguese were released with an arXiv paper in May.",
     "year": "2023"
    },
    {
     "title": "glue-ptpt and parlamento-pt datasets published",
     "note": "Datasets released on the PORTULAN Hugging Face organisation on 12 May.",
     "year": "2023"
    },
    {
     "title": "Albertina family extended",
     "note": "SIGUL 2024 paper added 1.5B and 100M parameter models.",
     "year": "2024"
    },
    {
     "title": "extraglue datasets released",
     "note": "extraglue-instruct appeared in March and extraglue in December.",
     "year": "2024"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "PORTULAN/albertina-900m-portuguese-ptbr-encoder-brwac",
     "url": "https://huggingface.co/PORTULAN/albertina-900m-portuguese-ptbr-encoder-brwac",
     "kind": "model",
     "note": "Model with 39 likes on the Hub."
    },
    {
     "name": "PORTULAN/albertina-900m-portuguese-ptpt-encoder",
     "url": "https://huggingface.co/PORTULAN/albertina-900m-portuguese-ptpt-encoder",
     "kind": "model",
     "note": "Model with 34 likes on the Hub."
    },
    {
     "name": "PORTULAN/extraglue",
     "url": "https://huggingface.co/datasets/PORTULAN/extraglue",
     "kind": "dataset",
     "note": "Dataset with 7 likes on the Hub."
    },
    {
     "name": "PORTULAN/portuguese-llm-leaderboard",
     "url": "https://huggingface.co/spaces/PORTULAN/portuguese-llm-leaderboard",
     "kind": "space",
     "note": "Space with 0 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Advancing Neural Encoding of Portuguese with Transformer Albertina PT-*",
     "url": "https://arxiv.org/abs/2305.06721",
     "year": "2023"
    },
    {
     "title": "Advancing Generative AI for Portuguese with Open Decoder Gervásio PT*",
     "url": "https://arxiv.org/abs/2402.18766",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Albertina 900M PT-PT",
     "url": "https://huggingface.co/PORTULAN/albertina-900m-portuguese-ptpt-encoder",
     "note": "Portuguese encoder"
    },
    {
     "name": "Albertina PT-BR",
     "url": "https://huggingface.co/PORTULAN/albertina-900m-portuguese-ptbr-encoder-brwac",
     "note": "Brazilian variant"
    },
    {
     "name": "Gervasio 7B",
     "url": "https://huggingface.co/PORTULAN/gervasio-7b-portuguese-ptbr-decoder",
     "note": "Portuguese decoder"
    },
    {
     "name": "extraglue",
     "url": "https://huggingface.co/datasets/PORTULAN/extraglue",
     "note": "Portuguese GLUE-style benchmark"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face under the PORTULAN organisation.",
   "key_people": [
    {
     "name": "Antonio Branco",
     "role": "Group leader",
     "url": "https://huggingface.co/PORTULAN",
     "note": null
    },
    {
     "name": "Rodrigo Santos",
     "role": "Researcher",
     "url": "https://huggingface.co/PORTULAN",
     "note": null
    },
    {
     "name": "Luis Gomes",
     "role": "Researcher",
     "url": "https://huggingface.co/PORTULAN",
     "note": null
    },
    {
     "name": "Tomas Osorio",
     "role": "Researcher",
     "url": "https://huggingface.co/PORTULAN",
     "note": null
    },
    {
     "name": "Joao Ricardo Silva",
     "role": "Researcher",
     "url": "https://huggingface.co/PORTULAN",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "http://nlx.di.fc.ul.pt",
    "https://huggingface.co/PORTULAN",
    "https://arxiv.org/abs/2305.06721"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "nlx-group",
   "mlat": 38.75766,
   "mlon": -9.1393,
   "logo": true
  },
  {
   "orig_id": "nomic-ai",
   "name": "Nomic AI",
   "alias": null,
   "full_name": "Nomic AI",
   "type": "startup",
   "founded": 2022,
   "country": "United States",
   "city": "New York",
   "lat": 40.7128,
   "lon": -74.006,
   "website": "https://www.nomic.ai",
   "huggingface": "https://huggingface.co/nomic-ai",
   "github": "https://github.com/nomic-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "New York company behind open embedding models and GPT4All, a local LLM app.",
   "tagline": "New York company that began with open embeddings and the GPT4All local-LLM app and now builds AI agents for architecture, engineering and construction.",
   "about": [
    "Nomic was founded in 2022 in New York. Its early work focused on open research and tools for understanding the data behind AI models, including the Atlas data explorer, GPT4All and the Nomic Embed models, which were released with open weights, open data and training code.",
    "Working with engineering firms led the company to focus on drawings, specifications, codes and project data. Its platform now offers agents and an API for architecture, engineering and construction, along with a benchmark called AEC-Bench.",
    "Coatue led a 17 million dollar Series A in 2023, and Aurecon and Arcadis joined as strategic investors in 2026."
   ],
   "mission": "Nomic says AI should give experts more time to be experts, taking on repetitive work while accountable professionals stay in control of results.",
   "notice": null,
   "known_for": [
    "Nomic Embed open embedding models",
    "GPT4All local language model app",
    "Atlas data exploration"
   ],
   "focus": [
    "embeddings",
    "local inference",
    "data visualisation"
   ],
   "contributions": [
    {
     "title": "Atlas",
     "note": "A browser tool for exploring large unstructured datasets.",
     "year": "2022"
    },
    {
     "title": "GPT4All",
     "note": "An app for running language models locally.",
     "year": "2023"
    },
    {
     "title": "Series A",
     "note": "Coatue led a 17 million dollar Series A round.",
     "year": "2023"
    },
    {
     "title": "Nomic Embed",
     "note": "Text and vision embedding models released with open weights, open data and training code.",
     "year": "2024"
    },
    {
     "title": "AEC-Bench",
     "note": "A benchmark for agentic systems in architecture, engineering and construction.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "gpt4all",
     "url": "https://github.com/nomic-ai/gpt4all",
     "note": "Run local LLMs on any device; most-starred repo."
    },
    {
     "name": "nomic",
     "url": "https://github.com/nomic-ai/nomic",
     "note": "Nomic developer API SDK."
    },
    {
     "name": "gpt4all-chat",
     "url": "https://github.com/nomic-ai/gpt4all-chat",
     "note": "Early GPT4All-J chat client."
    },
    {
     "name": "deepscatter",
     "url": "https://github.com/nomic-ai/deepscatter",
     "note": "Browser scatterplots at scale."
    }
   ],
   "hf_highlights": [
    {
     "name": "nomic-embed-text-v1.5",
     "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5",
     "kind": "model",
     "note": "Most downloaded embedding model."
    },
    {
     "name": "nomic-embed-text-v1",
     "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v1",
     "kind": "model",
     "note": "First open Nomic Embed."
    },
    {
     "name": "nomic-embed-text-v2-moe",
     "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe",
     "kind": "model",
     "note": "Multilingual MoE embedder."
    },
    {
     "name": "nomic-embed-unsupervised-data",
     "url": "https://huggingface.co/datasets/nomic-ai/nomic-embed-unsupervised-data",
     "kind": "dataset",
     "note": "Contrastive pretraining data."
    }
   ],
   "papers": [
    {
     "title": "Nomic Embed: Training a Reproducible Long Context Text Embedder",
     "url": "https://arxiv.org/abs/2402.01613",
     "year": "2024"
    },
    {
     "title": "Nomic Embed Vision: Expanding the Latent Space",
     "url": "https://arxiv.org/abs/2406.18587",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "gpt4all",
     "url": "https://github.com/nomic-ai/gpt4all",
     "note": "Local LLM app"
    },
    {
     "name": "nomic-embed-text-v1.5",
     "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5",
     "note": "Embedding model"
    }
   ],
   "openness": "Open weights, data and training code for embeddings.",
   "key_people": [
    {
     "name": "Andriy Mulyar",
     "role": "Founder and CEO",
     "url": "https://www.nomic.ai/about",
     "note": null
    },
    {
     "name": "Ben Schmidt",
     "role": "CTO",
     "url": "https://benschmidt.org",
     "note": null
    },
    {
     "name": "Juliet Fern",
     "role": "COO",
     "url": "https://www.nomic.ai/about",
     "note": null
    },
    {
     "name": "Chris Garrison",
     "role": "VP of Revenue",
     "url": "https://www.nomic.ai/about",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.nomic.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.nomic.ai",
    "https://huggingface.co/nomic-ai",
    "https://github.com/nomic-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "nomic-ai",
   "mlat": 40.75803,
   "mlon": -73.96265,
   "logo": true
  },
  {
   "orig_id": "norwai",
   "name": "NorwAI",
   "alias": "NTNU",
   "full_name": "NorwAI (NTNU)",
   "type": "institute",
   "founded": 2020,
   "country": "Norway",
   "city": "Trondheim",
   "lat": 63.4305,
   "lon": 10.3951,
   "website": "https://www.ntnu.edu/norwai",
   "huggingface": "https://huggingface.co/NorwAI",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Norwegian Research Center for AI Innovation at NTNU, publishing open-weight Norwegian language models.",
   "tagline": "NTNU's national AI innovation centre, publishing Norwegian large language models for research and experimentation.",
   "about": [
    "NorwAI is the Norwegian Research Center for AI Innovation at NTNU in Trondheim, started in October 2020 and directed by Professor Jon Atle Gulla. It combines academic research with industry partners that apply AI.",
    "In May 2024 the centre launched a suite of Norwegian large language models, built by adapting open architectures such as Mistral and Llama to Norwegian data and trained on NTNU's Idun cluster.",
    "Models are hosted on Hugging Face and can be downloaded free of charge for research and experiments. The centre states it has neither the mandate nor the resources to serve them as a permanent public API."
   ],
   "mission": "NorwAI aims to build Norwegian-language AI capacity and to carry research results into Norwegian industry and public services.",
   "notice": null,
   "known_for": [
    "Norwegian large language models",
    "NTNU Idun-trained NorLLM suite",
    "Industry-linked AI research centre"
   ],
   "focus": [
    "Norwegian LLMs"
   ],
   "contributions": [
    {
     "title": "NorwAI starts operations",
     "note": "The centre began work in October 2020.",
     "year": "2020"
    },
    {
     "title": "Large Norwegian GPT model announced",
     "note": "In January 2023 NorwAI described plans for a 20-billion-parameter Norwegian model trained on NTNU's Idun cluster, with partners including Schibsted and DnB.",
     "year": "2023"
    },
    {
     "title": "Norwegian LLM suite launched",
     "note": "Models adapted from open architectures such as Mistral and Llama were published on Hugging Face, beginning with NorwAI-Llama2-7B.",
     "year": "2024"
    },
    {
     "title": "Next model planned from Magistral",
     "note": "In September 2025 the centre said its next, larger model would start from Mistral's 24B Magistral and avoid copyrighted material.",
     "year": "2025"
    },
    {
     "title": "Technical report on NorwAI LLMs",
     "note": "Gulla, Liu and Zhang documented the model suite on arXiv in January 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "NorwAI/NorwAI-Magistral-24B-reasoning",
     "url": "https://huggingface.co/NorwAI/NorwAI-Magistral-24B-reasoning",
     "kind": "model",
     "note": "Model with 16 likes on the Hub."
    },
    {
     "name": "NorwAI/NorwAI-Mistral-7B-instruct",
     "url": "https://huggingface.co/NorwAI/NorwAI-Mistral-7B-instruct",
     "kind": "model",
     "note": "Model with 12 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "NorwAI's Large Language Models: Technical Report",
     "url": "https://arxiv.org/abs/2601.03034",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "NorwAI-Mistral-7B",
     "url": "https://huggingface.co/NorwAI/NorwAI-Mistral-7B",
     "note": "Norwegian adaptation of Mistral 7B"
    },
    {
     "name": "NorwAI-Mistral-7B-instruct",
     "url": "https://huggingface.co/NorwAI/NorwAI-Mistral-7B-instruct",
     "note": "Instruction-tuned variant"
    },
    {
     "name": "NorwAI-Llama2-7B",
     "url": "https://huggingface.co/NorwAI/NorwAI-Llama2-7B",
     "note": "Norwegian Llama 2 model"
    },
    {
     "name": "NorwAI-Magistral-24B-reasoning",
     "url": "https://huggingface.co/NorwAI/NorwAI-Magistral-24B-reasoning",
     "note": "24B Norwegian reasoning model"
    }
   ],
   "openness": "Model weights are public on the NorwAI Hugging Face organisation.",
   "key_people": [
    {
     "name": "Jon Atle Gulla",
     "role": "Director, NorwAI; Professor, NTNU",
     "url": "https://www.ntnu.edu/employees/jon.atle.gulla",
     "note": "Directs the centre and is lead author of its 2026 technical report on the NorLLM models."
    },
    {
     "name": "Peng Liu",
     "role": "Postdoctoral fellow (language models), NTNU",
     "url": "https://www.ntnu.edu/employees/penl",
     "note": "Worked on language models and conversational systems at the centre from 2020; co-author of the technical report."
    },
    {
     "name": "Lemei Zhang",
     "role": "Postdoctoral fellow, NTNU",
     "url": "https://www.ntnu.edu/employees/lemei.zhang",
     "note": "Researcher in streaming data analytics; co-author of the technical report."
    },
    {
     "name": "Lars Bungum",
     "role": "Postdoctoral fellow, NTNU",
     "url": "https://www.ntnu.edu/employees/lars.bungum",
     "note": "Works on evaluating trustworthy and efficient language models."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.ntnu.edu/norwai",
    "https://huggingface.co/NorwAI"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "norwai",
   "logo": true
  },
  {
   "orig_id": "nous-research",
   "name": "Nous Research",
   "alias": null,
   "full_name": "Nous Research",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "New York",
   "lat": 40.7128,
   "lon": -74.006,
   "website": "https://nousresearch.com",
   "huggingface": "https://huggingface.co/NousResearch",
   "github": "https://github.com/NousResearch",
   "x": "https://x.com/nousresearch",
   "discord": "https://discord.gg/nousresearch",
   "linkedin": null,
   "contact_email": "info@nousresearch.com",
   "summary": "Nous Research is an AI lab releasing open models such as the Hermes series, datasets and agent software.",
   "tagline": "American open AI lab behind the Hermes models, the Hermes Agent harness and decentralized training work such as DisTrO and Psyche.",
   "about": [
    "Nous Research began in 2022 as a group of volunteers who met on Discord, GitHub and Twitter, and became a company in 2023. Reported co-founders include Jeffrey Quesnelle, Karan Malhotra, Teknium and Shivani Mitra, and Bowen Peng is also described as a co-founder.",
    "The group first fine-tuned open models into the Hermes series and published YaRN, a context-extension method. It later released the DisTrO optimizer and the Psyche network for training over the public internet, and in February 2026 launched Hermes Agent, an open-source agent harness.",
    "Funding includes a 50 million dollar Series A led by Paradigm in April 2025 and a 90 million dollar round announced in October 2026."
   ],
   "mission": "Nous states its mission as creating and spreading open access to intelligence, so that powerful AI is in the hands of the many.",
   "notice": null,
   "known_for": [
    "Hermes open-weight models",
    "Hermes Agent open-source harness",
    "DisTrO and Psyche distributed training",
    "YaRN context extension"
   ],
   "focus": [
    "language models",
    "datasets",
    "agents",
    "distributed"
   ],
   "contributions": [
    {
     "title": "Seed round",
     "note": "Nous closed a 5.2 million dollar seed round in January 2024.",
     "year": "2024"
    },
    {
     "title": "Hermes 3 and DisTrO",
     "note": "Hermes 3 on Llama 3.1 and the DisTrO optimizer were published in August 2024.",
     "year": "2024"
    },
    {
     "title": "Series A and Psyche",
     "note": "A 50 million dollar round led by Paradigm accompanied a test network for the Psyche training system in April 2025.",
     "year": "2025"
    },
    {
     "title": "Hermes 4",
     "note": "The Hermes 4 models followed in August 2025.",
     "year": "2025"
    },
    {
     "title": "Hermes Agent",
     "note": "An open-source agent released in February 2026 that reached about 214,000 GitHub stars by July.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "hermes-agent",
     "url": "https://github.com/NousResearch/hermes-agent",
     "note": "Agent framework that grows with the user; the org's flagship repo."
    },
    {
     "name": "hermes-agent-self-evolution",
     "url": "https://github.com/NousResearch/hermes-agent-self-evolution",
     "note": "Evolutionary self-improvement tooling for Hermes Agent."
    },
    {
     "name": "hermes-paperclip-adapter",
     "url": "https://github.com/NousResearch/hermes-paperclip-adapter",
     "note": "Adapter for running Hermes as a managed agent."
    },
    {
     "name": "Hermes-Function-Calling",
     "url": "https://github.com/NousResearch/Hermes-Function-Calling",
     "note": "Reference code and prompts for Hermes function calling."
    }
   ],
   "hf_highlights": [
    {
     "name": "Hermes-3-Llama-3.1-8B",
     "url": "https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B",
     "kind": "model",
     "note": "Hermes 3 at 8B, the most downloaded Nous-authored model."
    },
    {
     "name": "hermes-function-calling-v1",
     "url": "https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1",
     "kind": "dataset",
     "note": "Function-calling and JSON-mode training data."
    },
    {
     "name": "json-mode-eval",
     "url": "https://huggingface.co/datasets/NousResearch/json-mode-eval",
     "kind": "dataset",
     "note": "Small JSON-mode evaluation dataset."
    }
   ],
   "papers": [
    {
     "title": "Hermes 3 Technical Report",
     "url": "https://arxiv.org/abs/2408.11857",
     "year": "2024"
    },
    {
     "title": "YaRN: Efficient Context Window Extension of Large Language Models",
     "url": "https://arxiv.org/abs/2309.00071",
     "year": "2023"
    },
    {
     "title": "DeMo: Decoupled Momentum Optimization",
     "url": "https://arxiv.org/abs/2411.19870",
     "year": "2024"
    },
    {
     "title": "Hermes 4 Technical Report",
     "url": "https://arxiv.org/abs/2508.18255",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Hermes-3-Llama-3.1-8B",
     "url": "https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B",
     "note": "Open model"
    },
    {
     "name": "hermes-agent",
     "url": "https://github.com/NousResearch/hermes-agent",
     "note": "Agent framework"
    },
    {
     "name": "hermes-function-calling-v1",
     "url": "https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1",
     "note": "Dataset"
    }
   ],
   "openness": "Open weights and code; model licences follow their base models.",
   "key_people": [
    {
     "name": "Jeffrey Quesnelle",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Nous_Research",
     "note": null
    },
    {
     "name": "Karan Malhotra",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Nous_Research",
     "note": null
    },
    {
     "name": "Bowen Peng",
     "role": "Researcher, described as co-founder",
     "url": "https://en.wikipedia.org/wiki/Nous_Research",
     "note": null
    },
    {
     "name": "Dillon Rolnick",
     "role": "Chief executive as of October 2026",
     "url": "https://en.wikipedia.org/wiki/Nous_Research",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://nousresearch.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co/NousResearch",
    "https://github.com/NousResearch",
    "https://arxiv.org/abs/2408.11857"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "nous-research",
   "mlat": 40.75803,
   "mlon": -74.04935,
   "logo": true
  },
  {
   "orig_id": "novasky",
   "name": "NovaSky",
   "alias": "Berkeley Sky Lab",
   "full_name": "NovaSky (Berkeley Sky Lab)",
   "type": "academic",
   "founded": 2025,
   "country": "United States",
   "city": "Berkeley",
   "lat": 37.8715,
   "lon": -122.273,
   "website": "https://novasky-ai.github.io",
   "huggingface": "https://huggingface.co/NovaSky-AI",
   "github": "https://github.com/NovaSky-AI",
   "x": null,
   "discord": "https://discord.gg/kexQXy5yA3",
   "linkedin": null,
   "contact_email": null,
   "summary": "UC Berkeley Sky Computing Lab group releasing open reasoning models and RL libraries such as Sky-T1 and SkyRL.",
   "tagline": "Berkeley Sky Computing Lab group releasing open reasoning models, training recipes and the SkyRL reinforcement-learning library for LLM agents.",
   "about": [
    "NovaSky is a team within the Sky Computing Lab at UC Berkeley. Its GitHub organization was created in January 2025, the month it published Sky-T1, a reasoning model it said could be trained for under 450 dollars.",
    "The group releases weights, data and training code together. After Sky-T1 it published work on cutting reasoning costs, test-time scaling for code generation and the SkyRL framework, which grew from SkyRL-v0 for long-horizon software agents into a modular RL library for training LLMs.",
    "Its repositories, including SkyThought and SkyRL, are released under Apache 2.0, and its audience is researchers and developers training reasoning and agentic models."
   ],
   "mission": "NovaSky aims to make post-training of reasoning and agentic models reproducible and affordable by sharing models, data and code openly.",
   "notice": null,
   "known_for": [
    "Sky-T1 reasoning model trained for under 450 dollars",
    "SkyRL reinforcement-learning library",
    "SkyThought open recipes",
    "Text-to-SQL and software-agent RL results"
   ],
   "focus": [
    "reasoning",
    "reinforcement learning",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Sky-T1-32B-Preview",
     "note": "An open 32-billion-parameter reasoning model, with data and code, reported to cost under 450 dollars to train.",
     "year": "2025"
    },
    {
     "title": "Reasoning cost reduction",
     "note": "A January 2025 post showed how to cut reasoning costs by about half without sacrificing accuracy.",
     "year": "2025"
    },
    {
     "title": "SkyRL-v0",
     "note": "A May 2025 release for training long-horizon real-world agents with reinforcement learning.",
     "year": "2025"
    },
    {
     "title": "SkyRL-SQL",
     "note": "A May 2025 recipe for data-efficient multi-turn reinforcement learning on text-to-SQL.",
     "year": "2025"
    },
    {
     "title": "SkyRL-v0.1",
     "note": "A June 2025 release evolving SkyRL into a modular, performant RL framework for LLMs.",
     "year": "2025"
    },
    {
     "title": "SA-SWE-32B",
     "note": "A 32-billion-parameter software-engineering model published in October 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "SkyThought",
     "url": "https://github.com/NovaSky-AI/SkyThought",
     "note": "Sky-T1 training recipes."
    },
    {
     "name": "SkyRL",
     "url": "https://github.com/NovaSky-AI/SkyRL",
     "note": "Modular full-stack RL library."
    },
    {
     "name": "SkyRL-OpenHands",
     "url": "https://github.com/NovaSky-AI/SkyRL-OpenHands",
     "note": "SkyRL with OpenHands agents."
    }
   ],
   "hf_highlights": [
    {
     "name": "SWE-Gym-OpenHands-7B-Agent",
     "url": "https://huggingface.co/NovaSky-AI/SWE-Gym-OpenHands-7B-Agent",
     "kind": "model",
     "note": "SWE agent model."
    },
    {
     "name": "Sky-T1-32B-Preview",
     "url": "https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview",
     "kind": "model",
     "note": "Reasoning model."
    },
    {
     "name": "Sky-T1_data_17k",
     "url": "https://huggingface.co/datasets/NovaSky-AI/Sky-T1_data_17k",
     "kind": "dataset",
     "note": "Sky-T1 training data."
    }
   ],
   "papers": [
    {
     "title": "LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!",
     "url": "https://arxiv.org/abs/2502.07374",
     "year": "2025"
    },
    {
     "title": "S*: Test Time Scaling for Code Generation",
     "url": "https://arxiv.org/abs/2502.14382",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Sky-T1-32B-Preview",
     "url": "https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview",
     "note": "Reasoning model"
    },
    {
     "name": "SkyRL",
     "url": "https://github.com/NovaSky-AI/SkyRL",
     "note": "RL library"
    }
   ],
   "openness": "Open weights, data and code.",
   "key_people": [
    {
     "name": "Shiyi Cao",
     "role": "Public org member",
     "url": "https://shiyicao.com",
     "note": null
    },
    {
     "name": "Charlie Ruan",
     "role": "Public org member",
     "url": "https://www.charlieruan.com",
     "note": null
    },
    {
     "name": "Shu Liu",
     "role": "Public org member",
     "url": "https://shulynnliu.com/",
     "note": null
    },
    {
     "name": "Sumanth R Hegde",
     "role": "Public org member",
     "url": "https://sumanthrh.com",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://novasky-ai.github.io); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://novasky-ai.github.io",
    "https://huggingface.co/NovaSky-AI",
    "https://github.com/NovaSky-AI"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "novasky",
   "mlat": 37.85423,
   "mlon": -122.34035,
   "logo": true
  },
  {
   "orig_id": "nvidia",
   "name": "NVIDIA",
   "alias": null,
   "full_name": "NVIDIA",
   "type": "corporate",
   "founded": 1993,
   "country": "United States",
   "city": "Santa Clara",
   "lat": 37.3541,
   "lon": -121.9552,
   "website": "https://www.nvidia.com",
   "huggingface": "https://huggingface.co/nvidia",
   "github": "https://github.com/NVIDIA",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "NVIDIA is a large public company best known for GPUs, and it also publishes a broad set of open models, datasets and training software. Its open releases include the Nemotron language models, Cosmos world models, GR00T robotics models, and the NeMo speech toolkit with Parakeet and Canary models.",
   "tagline": "A GPU maker that also releases open models, datasets and tooling, from Nemotron language models to Cosmos world models and speech ASR.",
   "about": [
    "NVIDIA was founded in 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem and is headquartered in Santa Clara, California. It is primarily a hardware and software platform company, and its open-model work sits alongside that business.",
    "Through its Hugging Face organization and GitHub, NVIDIA publishes model checkpoints, large synthetic and curated datasets, and training and inference software. Main families include Nemotron for language and reasoning, Cosmos for physical-AI world models, Isaac GR00T for humanoid robots, and Parakeet, Canary and Magpie for speech.",
    "Research is published as technical reports on arXiv, often with long author lists, and many releases come with training recipes and data. NVIDIA says its Nemotron effort includes open weights, open datasets and published methods."
   ],
   "mission": "NVIDIA presents its open releases as a way to let developers and organizations build, customize and deploy AI on its platform, including sovereign and domain-specific systems. It describes Nemotron as open technology covering models, data and recipes.",
   "notice": "NVIDIA is a large public company, listed here for its extensive open model and dataset releases. Not all of its products or models are open, and licence terms differ by model.",
   "known_for": [
    "Nemotron open language models and datasets",
    "Cosmos world foundation models for physical AI",
    "GR00T open humanoid robot models",
    "NeMo toolkit with Parakeet and Canary speech models"
   ],
   "focus": [
    "language models",
    "datasets",
    "speech",
    "robotics",
    "vision"
   ],
   "contributions": [
    {
     "title": "Megatron-LM",
     "note": "Open research code for training large transformer models at scale, published on NVIDIA's GitHub.",
     "year": "2019"
    },
    {
     "title": "Canary-1B",
     "note": "Multilingual speech recognition and translation model from the NeMo family, released under CC-BY-NC-4.0.",
     "year": "2024"
    },
    {
     "title": "Nemotron-4 340B",
     "note": "Family of 340B-parameter models with a technical report; released under the NVIDIA Open Model License.",
     "year": "2024"
    },
    {
     "title": "Cosmos platform",
     "note": "World foundation model platform for physical AI, described in a January 2025 technical report.",
     "year": "2025"
    },
    {
     "title": "GR00T N1",
     "note": "Open foundation model for generalist humanoid robots, announced in March 2025.",
     "year": "2025"
    },
    {
     "title": "Nemotron-H and Nemotron 3",
     "note": "Hybrid Mamba-Transformer models, followed by the Nemotron 3 family of Nano, Super and Ultra announced in December 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "NVIDIA-NeMo/NeMo",
     "url": "https://github.com/NVIDIA-NeMo/NeMo",
     "note": "Open-source toolkit for speech models (ASR, TTS, speech LLMs), Apache-2.0."
    },
    {
     "name": "NVIDIA/Megatron-LM",
     "url": "https://github.com/NVIDIA/Megatron-LM",
     "note": "Research code for training transformer models at scale."
    },
    {
     "name": "NVIDIA/cosmos",
     "url": "https://github.com/NVIDIA/cosmos",
     "note": "Open platform of world models, datasets and tools for physical AI."
    },
    {
     "name": "NVIDIA-NeMo/Nemotron",
     "url": "https://github.com/NVIDIA-NeMo/Nemotron",
     "note": "Primary Nemotron repository, listed as Apache-2.0 on NVIDIA's model info page."
    }
   ],
   "hf_highlights": [
    {
     "name": "nvidia/parakeet-tdt-0.6b-v2",
     "url": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2",
     "kind": "model",
     "note": "600M-parameter English speech-to-text model under CC-BY-4.0."
    },
    {
     "name": "nvidia/canary-1b",
     "url": "https://huggingface.co/nvidia/canary-1b",
     "kind": "model",
     "note": "Multilingual ASR and translation model for English, German, French and Spanish (CC-BY-NC-4.0)."
    },
    {
     "name": "nvidia/AeroZoo",
     "url": "https://huggingface.co/datasets/nvidia/AeroZoo",
     "kind": "dataset",
     "note": "Recently updated dataset on the NVIDIA organization page."
    },
    {
     "name": "nvidia/nemotron-transcribe",
     "url": "https://huggingface.co/spaces/nvidia/nemotron-transcribe",
     "kind": "space",
     "note": "Demo space for NVIDIA's Nemotron multilingual transcription models."
    }
   ],
   "papers": [
    {
     "title": "Nemotron-4 340B Technical Report",
     "url": "https://arxiv.org/abs/2406.11704",
     "year": "2024"
    },
    {
     "title": "Cosmos World Foundation Model Platform for Physical AI",
     "url": "https://arxiv.org/abs/2501.03575",
     "year": "2025"
    },
    {
     "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots",
     "url": "https://arxiv.org/abs/2503.14734",
     "year": "2025"
    },
    {
     "title": "Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models",
     "url": "https://arxiv.org/abs/2504.03624",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Nemotron 3 (Nano, Super, Ultra)",
     "url": "https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/",
     "note": "Hybrid mixture-of-experts language models announced December 2025."
    },
    {
     "name": "Cosmos",
     "url": "https://github.com/NVIDIA/cosmos",
     "note": "Open world models, datasets and tools for physical AI."
    },
    {
     "name": "Isaac GR00T N1",
     "url": "https://arxiv.org/abs/2503.14734",
     "note": "Open humanoid robot foundation model."
    },
    {
     "name": "Parakeet and Canary",
     "url": "https://github.com/NVIDIA-NeMo/NeMo",
     "note": "Speech recognition models built with the NeMo toolkit."
    }
   ],
   "openness": "Many NVIDIA models ship as open weights, with datasets and code on Hugging Face and GitHub, but licences vary: the NVIDIA Open Model License covers Nemotron-4 and the Nemotron 3 Nano and Super models, newer Nemotron 3 Ultra and 3.5 models are reported under OpenMDW-1.1, NeMo code is Apache-2.0, and speech models range from CC-BY-4.0 (Parakeet) to CC-BY-NC-4.0 (Canary-1B).",
   "key_people": [
    {
     "name": "Bryan Catanzaro",
     "role": "Vice President, Applied Deep Learning Research; leads the Nemotron effort",
     "url": "https://blogs.nvidia.com/blog/nemotron-open-source-ai",
     "note": "Author of NVIDIA's blog post presenting Nemotron as open technology."
    },
    {
     "name": "Jensen Huang",
     "role": "Co-founder",
     "url": "https://en.wikipedia.org/wiki/Jensen_Huang",
     "note": "Co-founded NVIDIA in 1993; Wikipedia's Nvidia article currently reports a change in his CEO role."
    },
    {
     "name": "Linxi Fan",
     "role": "Project lead, GR00T N1",
     "url": "https://arxiv.org/abs/2503.14734",
     "note": "Known as Jim Fan; named as project lead on the GR00T N1 paper."
    },
    {
     "name": "Yuke Zhu",
     "role": "Project lead, GR00T N1",
     "url": "https://arxiv.org/abs/2503.14734",
     "note": "Named as project lead on the GR00T N1 paper."
    },
    {
     "name": "Sanja Fidler",
     "role": "Researcher, Cosmos",
     "url": "https://arxiv.org/abs/2501.03575",
     "note": "Listed among the Cosmos platform paper authors."
    }
   ],
   "discord_note": "No official NVIDIA Discord for open models was verified.",
   "contact_note": "No general contact address was verified; use NVIDIA's Hugging Face and GitHub pages for model and repository issues.",
   "sources": [
    "https://en.wikipedia.org/wiki/Nvidia",
    "https://huggingface.co/nvidia",
    "https://github.com/NVIDIA",
    "https://blogs.nvidia.com/blog/nemotron-open-source-ai",
    "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2",
    "https://huggingface.co/nvidia/canary-1b",
    "https://github.com/NVIDIA-NeMo/NeMo",
    "https://arxiv.org/abs/2406.11704"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "nvidia",
   "logo": true
  },
  {
   "orig_id": "nxai",
   "name": "NXAI",
   "alias": null,
   "full_name": "NXAI",
   "type": "startup",
   "founded": 2024,
   "country": "Austria",
   "city": "Linz",
   "lat": 48.3069,
   "lon": 14.2858,
   "website": "https://www.nx-ai.com",
   "huggingface": "https://huggingface.co/NX-AI",
   "github": "https://github.com/NX-AI",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/nxai",
   "contact_email": "contact@nx-ai.com",
   "summary": "Linz-based company commercialising xLSTM research from the JKU Linz group. It publishes xLSTM models, TiRex time-series forecasters and kernels.",
   "tagline": "Linz company commercializing Sepp Hochreiter's xLSTM research, publishing xLSTM models, TiRex time-series forecasters and fast recurrent kernels.",
   "about": [
    "NXAI was set up in early 2024 in Linz, Austria, by Sepp Hochreiter, who leads the Institute for Machine Learning at JKU Linz, together with the company builder Netural X and PIERER Digital Holding. Hochreiter is Chief Scientist and Albert Ortig is CEO, according to press coverage.",
    "The company aims to turn research on sequence modeling into industrial applications in engineering, robotics, automation and process control. It publishes the xLSTM architecture code, FlashRNN kernels, Vision-LSTM and the TiRex family of zero-shot time-series forecasting models.",
    "Its GitHub organization dates from May 2024, and recent Hugging Face releases include TiRex-2 forecasting assets and KronLSTM language models."
   ],
   "mission": "NXAI states that its mission is to turn pioneering research into market-relevant industrial AI by building scalable, generalist foundation models.",
   "notice": null,
   "known_for": [
    "xLSTM recurrent language-model architecture",
    "TiRex zero-shot time-series forecasting",
    "FlashRNN kernels",
    "Industry-focused AI from Linz"
   ],
   "focus": [
    "xLSTM",
    "time series",
    "architectures"
   ],
   "contributions": [
    {
     "title": "xLSTM repository",
     "note": "Reference implementation of the xLSTM architecture published in May 2024.",
     "year": "2024"
    },
    {
     "title": "Vision-LSTM",
     "note": "An xLSTM-based vision backbone repository created in June 2024.",
     "year": "2024"
    },
    {
     "title": "FlashRNN",
     "note": "Optimized kernels for recurrent networks released in November 2024.",
     "year": "2024"
    },
    {
     "title": "TiRex",
     "note": "A zero-shot time-series forecasting model built on xLSTM, with code from May 2025.",
     "year": "2025"
    },
    {
     "title": "TiRex-2",
     "note": "An updated forecasting model whose code was published in July 2026.",
     "year": "2026"
    },
    {
     "title": "KronLSTM models",
     "note": "1.5-billion-parameter recurrent language models published on Hugging Face in October 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "xlstm",
     "url": "https://github.com/NX-AI/xlstm",
     "note": "Official repository of the xLSTM (2,218 stars)"
    },
    {
     "name": "vision-lstm",
     "url": "https://github.com/NX-AI/vision-lstm",
     "note": "xLSTM as a generic vision backbone (490 stars)"
    },
    {
     "name": "tirex",
     "url": "https://github.com/NX-AI/tirex",
     "note": "Zero-shot forecasting model for long and short horizons (308 stars)"
    },
    {
     "name": "flashrnn",
     "url": "https://github.com/NX-AI/flashrnn",
     "note": "Fast RNN kernels with I/O awareness (190 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "NX-AI/TiRex",
     "url": "https://huggingface.co/NX-AI/TiRex",
     "kind": "model",
     "note": "Zero-shot time-series forecasting model, the most downloaded NX-AI repository."
    },
    {
     "name": "NX-AI/TiRex-2",
     "url": "https://huggingface.co/NX-AI/TiRex-2",
     "kind": "model",
     "note": "Second generation of the TiRex forecaster."
    },
    {
     "name": "NX-AI/xLSTM-7b",
     "url": "https://huggingface.co/NX-AI/xLSTM-7b",
     "kind": "model",
     "note": "7B xLSTM language model weights."
    },
    {
     "name": "NX-AI/TiRex-demo",
     "url": "https://huggingface.co/spaces/NX-AI/TiRex-demo",
     "kind": "space",
     "note": "Interactive demo of TiRex forecasting."
    }
   ],
   "papers": [
    {
     "title": "xLSTM: Extended Long Short-Term Memory",
     "url": "https://arxiv.org/abs/2405.04517",
     "year": "2024"
    },
    {
     "title": "xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference",
     "url": "https://arxiv.org/abs/2503.13427",
     "year": "2025"
    },
    {
     "title": "TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning",
     "url": "https://arxiv.org/abs/2505.23719",
     "year": "2025"
    },
    {
     "title": "FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware",
     "url": "https://arxiv.org/abs/2412.07752",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "xLSTM code",
     "url": "https://github.com/NX-AI/xlstm",
     "note": "Official xLSTM repository"
    },
    {
     "name": "TiRex",
     "url": "https://huggingface.co/NX-AI/TiRex",
     "note": "Zero-shot time series forecasting model"
    },
    {
     "name": "xLSTM-7b",
     "url": "https://huggingface.co/NX-AI/xLSTM-7b",
     "note": "7B xLSTM language model"
    },
    {
     "name": "FlashRNN",
     "url": "https://github.com/NX-AI/flashrnn",
     "note": "Fast RNN kernels"
    }
   ],
   "openness": "The xLSTM code is Apache-2.0 and several models are public on Hugging Face.",
   "key_people": [
    {
     "name": "Sepp Hochreiter",
     "role": "Co-founder and Chief Scientist; head of the Institute for Machine Learning, JKU Linz",
     "url": "https://en.wikipedia.org/wiki/Sepp_Hochreiter",
     "note": "Inventor of the LSTM and senior author of the xLSTM and TiRex papers."
    },
    {
     "name": "Albert Ortig",
     "role": "CEO",
     "url": "https://www.netural.com/en/news/netural-ai-symposium-report-part-1",
     "note": "Leads the company, according to press coverage."
    },
    {
     "name": "Maximilian Beck",
     "role": "Lead author of xLSTM",
     "url": "https://arxiv.org/abs/2405.04517",
     "note": "First author of the xLSTM paper from the Hochreiter group."
    },
    {
     "name": "Andreas Auer",
     "role": "Lead author of TiRex",
     "url": "https://arxiv.org/abs/2505.23719",
     "note": "First author of the TiRex forecasting paper presented at NeurIPS 2025."
    }
   ],
   "discord_note": null,
   "contact_note": "The company site prints a general contact address.",
   "sources": [
    "https://www.nx-ai.com",
    "https://huggingface.co/NX-AI",
    "https://github.com/NX-AI",
    "https://arxiv.org/abs/2405.04517"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "nxai",
   "mlat": 48.34226,
   "mlon": 14.2858,
   "logo": true
  },
  {
   "orig_id": "oatml",
   "name": "OATML",
   "alias": "Oxford",
   "full_name": "OATML (Oxford)",
   "type": "academic",
   "founded": 2018,
   "country": "United Kingdom",
   "city": "Oxford",
   "lat": 51.752,
   "lon": -1.2577,
   "website": "https://oatml.cs.ox.ac.uk",
   "huggingface": null,
   "github": "https://github.com/OATML",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "oatml@cs.ox.ac.uk",
   "summary": "Oxford Applied and Theoretical Machine Learning group in the Department of Computer Science at the University of Oxford, working on uncertainty, Bayesian deep learning and safe ML, with research code released on GitHub.",
   "tagline": "Yarin Gal's Oxford group on Bayesian deep learning, uncertainty and safe machine learning, with open research code.",
   "about": [
    "The Oxford Applied and Theoretical Machine Learning group sits in the Department of Computer Science at the University of Oxford. Yarin Gal announced it in October 2018 and leads it as Associate Professor of Machine Learning.",
    "Its research covers Bayesian deep learning, uncertainty estimation and the safety of machine learning systems. Code accompanying its papers is released on GitHub.",
    "Former members include Sebastian Farquhar, Clare Lyle, Pascal Notin and Joost van Amersfoort."
   ],
   "mission": "The group works on principled machine learning methods that quantify uncertainty and make learning systems more reliable.",
   "notice": null,
   "known_for": [
    "Bayesian deep learning and uncertainty",
    "Yarin Gal's group",
    "Open research code on GitHub"
   ],
   "focus": [
    "uncertainty",
    "Bayesian deep learning"
   ],
   "contributions": [
    {
     "title": "OATML announced",
     "note": "Yarin Gal announced the group on 30 October.",
     "year": "2018"
    },
    {
     "title": "Group web presence at oatml.cs.ox.ac.uk",
     "note": "The group publishes members, publications and an application page.",
     "year": "2019"
    },
    {
     "title": "Member pages for alumni such as Aidan Gomez",
     "note": "Group site lists collaborators and affiliates.",
     "year": "2020"
    },
    {
     "title": "Group members span uncertainty and LLM safety",
     "note": "The 2026 member list covers postdocs and PhD students working on language models, safety and uncertainty under Yarin Gal.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "bdl-benchmarks",
     "url": "https://github.com/OATML/bdl-benchmarks",
     "note": "Bayesian Deep Learning Benchmarks (672 stars)"
    },
    {
     "name": "non-parametric-transformers",
     "url": "https://github.com/OATML/non-parametric-transformers",
     "note": "Code for \"Self-Attention Between Datapoints: Going Beyond Individual Input-Output Pairs in Deep Learning\" (418 stars)"
    },
    {
     "name": "RHO-Loss",
     "url": "https://github.com/OATML/RHO-Loss",
     "note": "Open-source repository. (217 stars)"
    },
    {
     "name": "oatomobile",
     "url": "https://github.com/OATML/oatomobile",
     "note": "A research framework for autonomous driving (199 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Self-Attention Between Datapoints: Going Beyond Individual Input-Output Pairs in Deep Learning",
     "url": "https://arxiv.org/abs/2106.02584",
     "year": "2021"
    },
    {
     "title": "Prioritized Training on Points that are Learnable, Worth Learning, and Not Yet Learnt",
     "url": "https://arxiv.org/abs/2206.07137",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "bdl-benchmarks",
     "url": "https://github.com/OATML/bdl-benchmarks",
     "note": "Benchmark suite for Bayesian deep learning"
    },
    {
     "name": "Non-parametric transformers",
     "url": "https://github.com/OATML/non-parametric-transformers",
     "note": "Code for attention between datapoints"
    },
    {
     "name": "RHO-Loss",
     "url": "https://github.com/OATML/RHO-Loss",
     "note": "Code for prioritised training on learnable points"
    },
    {
     "name": "OATomobile",
     "url": "https://github.com/OATML/oatomobile",
     "note": "Research framework for autonomous driving"
    }
   ],
   "openness": "Papers ship with public GitHub repositories; the group also hosts datasets and benchmarks on its site.",
   "key_people": [
    {
     "name": "Yarin Gal",
     "role": "Head of group, Associate Professor",
     "url": "https://cs.ox.ac.uk/people/yarin.gal/website",
     "note": null
    },
    {
     "name": "Aidan Gomez",
     "role": "Member",
     "url": "https://oatml.cs.ox.ac.uk/members/aidan_gomez/",
     "note": null
    },
    {
     "name": "Robin Ru",
     "role": "Member",
     "url": "https://www.cs.ox.ac.uk/oatml/members/robin_ru/",
     "note": null
    },
    {
     "name": "Pascal Notin",
     "role": "Associate member (Senior Research Fellow)",
     "url": "https://oatml.cs.ox.ac.uk/members/pascal_notin/",
     "note": "Former group member working on machine learning for biology."
    },
    {
     "name": "Ilia Shumailov",
     "role": "Associate member (Senior Research Fellow)",
     "url": "https://oatml.cs.ox.ac.uk/members/ilia_shumailov/",
     "note": "Associate member listed from 2022."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via address on the group website.",
   "sources": [
    "https://oatml.cs.ox.ac.uk",
    "https://github.com/OATML",
    "https://arxiv.org/abs/2106.02584"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "oatml",
   "mlat": 51.71664,
   "mlon": -1.2577,
   "logo": true
  },
  {
   "orig_id": "occiglot",
   "name": "Occiglot",
   "alias": null,
   "full_name": "Occiglot",
   "type": "collective",
   "founded": 2024,
   "country": "Germany",
   "city": "",
   "lat": 51.1657,
   "lon": 10.4515,
   "website": "https://occiglot.eu",
   "huggingface": "https://huggingface.co/occiglot",
   "github": "https://github.com/occiglot",
   "x": "https://x.com/occiglot",
   "discord": "https://discord.gg/wUpvYs4XvM",
   "linkedin": null,
   "contact_email": null,
   "summary": "Research collective developing open language models for European languages. It released the Occiglot-7B series and a European LLM leaderboard.",
   "tagline": "An informal European research collective that released Occiglot-7B models, multilingual web datasets and tokenizer analyses for European languages.",
   "about": [
    "Occiglot describes itself as an ongoing research collective and an informal community of motivated individuals building open multilingual language models for Europe. Its announcement in March 2024 framed the work as preserving linguistic diversity and AI sovereignty in Europe.",
    "The group released continually pre-trained Occiglot-7B models for German, French, Italian, Spanish and combined European language sets in March 2024, followed by the Occiglot-Fineweb dataset and German models built with DiscoResearch. It also published a tokenizer evaluation for European languages and contributed to the Community OSCAR web corpus.",
    "The page names contributors Manuel Brack, Malte Ostendorff and Pedro Ortiz Suarez, and invites collaborators through its Discord server."
   ],
   "mission": "Occiglot works to build transparent, open-source multilingual language models for European languages through an open community.",
   "notice": null,
   "known_for": [
    "Occiglot-7B multilingual European models",
    "Occiglot-Fineweb dataset",
    "European language tokenizer evaluation",
    "Community OSCAR web data"
   ],
   "focus": [
    "language models",
    "European languages",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Occiglot announcement",
     "note": "The collective published its mission post on 6 March 2024.",
     "year": "2024"
    },
    {
     "title": "Occiglot-7B model family",
     "note": "Language-specific and eu5 Occiglot-7B models appeared on Hugging Face in early March 2024.",
     "year": "2024"
    },
    {
     "title": "Tokenizer evaluation",
     "note": "An analysis of tokenizers on European languages published on 11 March 2024.",
     "year": "2024"
    },
    {
     "title": "Occiglot-Fineweb",
     "note": "A preliminary dataset of about 230 million cleaned documents in 10 languages.",
     "year": "2024"
    },
    {
     "title": "German Llama-3 models",
     "note": "Base, instruct and long-context German models built with DiscoResearch on 65 billion German tokens.",
     "year": "2024"
    },
    {
     "title": "Community OSCAR",
     "note": "A multilingual web dataset built from 40 monthly Common Crawl snapshots, announced in August 2024.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "euro-lm-evaluation-harness",
     "url": "https://github.com/occiglot/euro-lm-evaluation-harness",
     "note": "Evaluation harness for European language models (6 stars)"
    },
    {
     "name": "tech-report",
     "url": "https://github.com/occiglot/tech-report",
     "note": "Raw data and scripts for the technical report (6 stars)"
    },
    {
     "name": "curated-web-data",
     "url": "https://github.com/occiglot/curated-web-data",
     "note": "Collection of good and bad web domains for LLM pretraining data (4 stars)"
    },
    {
     "name": "occiglot",
     "url": "https://github.com/occiglot/occiglot",
     "note": "Main organisation repository (4 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "occiglot/occiglot-7b-it-en-instruct",
     "url": "https://huggingface.co/occiglot/occiglot-7b-it-en-instruct",
     "kind": "model",
     "note": "Italian-English instruction-tuned 7B model."
    },
    {
     "name": "occiglot/occiglot-7b-de-en-instruct",
     "url": "https://huggingface.co/occiglot/occiglot-7b-de-en-instruct",
     "kind": "model",
     "note": "German-English instruction-tuned 7B model."
    },
    {
     "name": "occiglot/tokenizer-wiki-bench",
     "url": "https://huggingface.co/datasets/occiglot/tokenizer-wiki-bench",
     "kind": "dataset",
     "note": "Tokenizer benchmark data from Wikipedia."
    },
    {
     "name": "occiglot/euro-llm-leaderboard",
     "url": "https://huggingface.co/spaces/occiglot/euro-llm-leaderboard",
     "kind": "space",
     "note": "European LLM leaderboard."
    }
   ],
   "papers": [
    {
     "title": "Occiglot at WMT24: European Open-source Large Language Models Evaluated on Translation",
     "url": "https://aclanthology.org/2024.wmt-1.23/",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "occiglot-7b-eu5",
     "url": "https://huggingface.co/occiglot/occiglot-7b-eu5",
     "note": "7B model for five European languages"
    },
    {
     "name": "occiglot-fineweb",
     "url": "https://huggingface.co/datasets/occiglot/occiglot-fineweb-v0.5",
     "note": "Multilingual web dataset"
    },
    {
     "name": "Euro LLM Leaderboard",
     "url": "https://huggingface.co/spaces/occiglot/euro-llm-leaderboard",
     "note": "Leaderboard for European LLMs"
    }
   ],
   "openness": "Models, datasets and evaluation code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Manuel Brack",
     "role": "Contributor",
     "url": "https://occiglot.eu",
     "note": null
    },
    {
     "name": "Malte Ostendorff",
     "role": "Contributor",
     "url": "https://occiglot.eu",
     "note": null
    },
    {
     "name": "Pedro Ortiz Suarez",
     "role": "Contributor",
     "url": "https://occiglot.eu",
     "note": null
    }
   ],
   "discord_note": "Discord of the informal collective for discussing multilingual LLM work; the site invites newcomers and collaborators to join.",
   "contact_note": "Occiglot's site invites contact and Discord joining from people working on multilingual LLMs.",
   "sources": [
    "https://occiglot.eu",
    "https://huggingface.co/occiglot",
    "https://github.com/occiglot"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "occiglot",
   "mlat": 51.20106,
   "mlon": 10.4515,
   "logo": true
  },
  {
   "orig_id": "open-climate-fix",
   "name": "Open Climate Fix",
   "alias": null,
   "full_name": "Open Climate Fix",
   "type": "nonprofit",
   "founded": 2019,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://openclimatefix.org",
   "huggingface": "https://huggingface.co/openclimatefix",
   "github": "https://github.com/openclimatefix",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/open-climate-fix",
   "contact_email": "info@openclimatefix.org",
   "summary": "UK non-profit applying machine learning to energy and climate problems such as solar forecasting. It releases open code, models and datasets.",
   "tagline": "UK organization applying machine learning to solar, wind and demand forecasting for the grid, publishing its code and data openly.",
   "about": [
    "Open Climate Fix was founded in 2019 by Dan Travers and Jack Kelly, and its GitHub organization describes its method as open science to mitigate climate change. The team combines expertise from energy, deep learning, design and product management.",
    "Its work includes a national solar forecast for Great Britain used daily by the National Energy System Operator, cloudcasting from satellite imagery, wind forecasting, mapping of rooftop solar and an open energy data advocacy program.",
    "The organization reports reductions in forecasting error and cost savings for grid operators and publishes models and code on GitHub and Hugging Face."
   ],
   "mission": "Open Climate Fix works toward a world powered by clean and affordable energy by delivering AI that makes renewable generation more predictable.",
   "notice": null,
   "known_for": [
    "GB National Solar Forecast",
    "Cloudcasting from satellite imagery",
    "Open energy data advocacy",
    "Open-source solar forecasting code"
   ],
   "focus": [
    "science",
    "climate",
    "forecasting",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Organization founded",
     "note": "Open Climate Fix was established by co-founders Dan Travers and Jack Kelly.",
     "year": "2019"
    },
    {
     "title": "GB National Solar Forecast",
     "note": "A solar PV generation forecast used daily by the National Energy System Operator and trading desks.",
     "year": "2024"
    },
    {
     "title": "Cloudcasting",
     "note": "Prediction of cloud movement from satellite imagery to improve solar forecasts.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "graph_weather",
     "url": "https://github.com/openclimatefix/graph_weather",
     "note": "Graph-based weather forecasting models (306 stars)"
    },
    {
     "name": "metnet",
     "url": "https://github.com/openclimatefix/metnet",
     "note": "PyTorch implementation of MetNet and MetNet-2 (305 stars)"
    },
    {
     "name": "skillful_nowcasting",
     "url": "https://github.com/openclimatefix/skillful_nowcasting",
     "note": "Implementation of DeepMind's Deep Generative Model of Radar (301 stars)"
    },
    {
     "name": "open-source-quartz-solar-forecast",
     "url": "https://github.com/openclimatefix/open-source-quartz-solar-forecast",
     "note": "Open-source site-level solar forecast (162 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "openclimatefix/dwd-icon-global",
     "url": "https://huggingface.co/datasets/openclimatefix/dwd-icon-global",
     "kind": "dataset",
     "note": "Global DWD ICON weather forecast data, the group's most downloaded dataset."
    },
    {
     "name": "openclimatefix/met-office-uk-deterministic-solar",
     "url": "https://huggingface.co/datasets/openclimatefix/met-office-uk-deterministic-solar",
     "kind": "dataset",
     "note": "Met Office UK deterministic forecast data for solar work."
    },
    {
     "name": "openclimatefix/dwd-icon-eu",
     "url": "https://huggingface.co/datasets/openclimatefix/dwd-icon-eu",
     "kind": "dataset",
     "note": "DWD ICON-EU regional forecast data."
    },
    {
     "name": "openclimatefix/dgmr",
     "url": "https://huggingface.co/openclimatefix/dgmr",
     "kind": "model",
     "note": "Radar nowcasting model."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Quartz Solar Forecast",
     "url": "https://github.com/openclimatefix/open-source-quartz-solar-forecast",
     "note": "Open-source site-level solar forecast"
    },
    {
     "name": "graph_weather",
     "url": "https://github.com/openclimatefix/graph_weather",
     "note": "Graph-based weather forecasting models"
    },
    {
     "name": "uk_pv dataset",
     "url": "https://huggingface.co/datasets/openclimatefix/uk_pv",
     "note": "UK solar PV dataset"
    }
   ],
   "openness": "Code is MIT-licensed and datasets are on Hugging Face.",
   "key_people": [
    {
     "name": "Dan Travers",
     "role": "Co-founder",
     "url": "https://openclimatefix.org/team",
     "note": null
    },
    {
     "name": "Jack Kelly",
     "role": "Co-founder",
     "url": "https://openclimatefix.org/team",
     "note": null
    },
    {
     "name": "Peter Dudfield",
     "role": "Head of Technology",
     "url": "https://openclimatefix.org/team",
     "note": null
    },
    {
     "name": "Heather Murphy",
     "role": "Commercial and Growth Strategy Lead",
     "url": "https://openclimatefix.org/team",
     "note": null
    },
    {
     "name": "Ben Nelmes",
     "role": "Open Data Advocacy Lead",
     "url": "https://openclimatefix.org/team",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Its site has a contact page with a form and a general information address.",
   "sources": [
    "https://openclimatefix.org",
    "https://huggingface.co/openclimatefix",
    "https://github.com/openclimatefix"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "open-climate-fix",
   "mlat": 51.56699,
   "mlon": -0.17391,
   "logo": true
  },
  {
   "orig_id": "openbmb",
   "name": "OpenBMB / ModelBest",
   "alias": null,
   "full_name": "OpenBMB / ModelBest",
   "type": "collective",
   "founded": 2021,
   "country": "China",
   "city": "Beijing",
   "lat": 39.9,
   "lon": 116.41,
   "website": "https://www.openbmb.cn",
   "huggingface": "https://huggingface.co/openbmb",
   "github": "https://github.com/OpenBMB",
   "x": "https://twitter.com/OpenBMB",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open lab for big model base working on foundation models and systems; known for the MiniCPM and MiniCPM-V model lines.",
   "tagline": "Open lab for big model base, linked to the Tsinghua NLP group's spin-out ModelBest, releasing the MiniCPM on-device models, VoxCPM speech and UltraRAG.",
   "about": [
    "OpenBMB, short for Open Lab for Big Model Base, describes its goal as building foundation models and systems toward AGI. It is the open-source community associated with ModelBest, a Beijing company founded in August 2022 as a spin-off from Tsinghua University's Natural Language Processing Laboratory.",
    "ModelBest was founded with Li Dahai as CEO and Liu Zhiyuan as Chief Scientist, according to press and encyclopedic sources. The community's flagship releases are the MiniCPM small language models and the MiniCPM-V multimodal models for local devices, along with VoxCPM for speech and the UltraRAG framework.",
    "ModelBest says the MiniCPM series has passed 24 million downloads, and its models appear in phones and cars."
   ],
   "mission": "OpenBMB aims to build foundation models and systems toward AGI, with efficient models that run on consumer devices.",
   "notice": null,
   "known_for": [
    "MiniCPM on-device language models",
    "MiniCPM-V multimodal models",
    "VoxCPM text-to-speech",
    "UltraRAG framework"
   ],
   "focus": [
    "language models",
    "multimodal",
    "efficient models",
    "Chinese"
   ],
   "contributions": [
    {
     "title": "ModelBest founded",
     "note": "The company was spun out of Tsinghua's NLP Laboratory in August 2022.",
     "year": "2022"
    },
    {
     "title": "Luca assistant",
     "note": "ModelBest launched its multimodal dialogue assistant Luca in November 2023.",
     "year": "2023"
    },
    {
     "title": "MiniCPM 4.0",
     "note": "Released in June 2025 alongside the CPM.cu inference framework.",
     "year": "2025"
    },
    {
     "title": "MiniCPM-o 4.5",
     "note": "A full-duplex multimodal model for video, audio and text input introduced in early 2026.",
     "year": "2026"
    },
    {
     "title": "MiniCPM5-2B",
     "note": "A two-billion-parameter model whose local deployment tests were published in November 2026.",
     "year": "2026"
    },
    {
     "title": "MiniCPM paper",
     "note": "The MiniCPM small language models were described in an April 2024 arXiv paper with 25 authors.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "VoxCPM",
     "url": "https://github.com/OpenBMB/VoxCPM",
     "note": "Tokenizer-free multilingual TTS."
    },
    {
     "name": "ChatDev",
     "url": "https://github.com/OpenBMB/ChatDev",
     "note": "Multi-agent software development."
    },
    {
     "name": "MiniCPM-V",
     "url": "https://github.com/OpenBMB/MiniCPM-V",
     "note": "On-device multimodal models."
    },
    {
     "name": "MiniCPM",
     "url": "https://github.com/OpenBMB/MiniCPM",
     "note": "On-device LLMs."
    }
   ],
   "hf_highlights": [
    {
     "name": "MiniCPM5-2B",
     "url": "https://huggingface.co/openbmb/MiniCPM5-2B",
     "kind": "model",
     "note": "Small on-device LLM."
    },
    {
     "name": "VoxCPM2",
     "url": "https://huggingface.co/openbmb/VoxCPM2",
     "kind": "model",
     "note": "Multilingual TTS."
    },
    {
     "name": "MiniCPM-V-4.7-35B-A3B",
     "url": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B",
     "kind": "model",
     "note": "MoE multimodal model."
    },
    {
     "name": "Ultra-FineWeb",
     "url": "https://huggingface.co/datasets/openbmb/Ultra-FineWeb",
     "kind": "dataset",
     "note": "Filtered web corpus."
    }
   ],
   "papers": [
    {
     "title": "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies",
     "url": "https://arxiv.org/abs/2404.06395",
     "year": "2024"
    },
    {
     "title": "MiniCPM-V: A GPT-4V Level MLLM on Your Phone",
     "url": "https://arxiv.org/abs/2408.01800",
     "year": "2024"
    },
    {
     "title": "ChatDev: Communicative Agents for Software Development",
     "url": "https://arxiv.org/abs/2307.07924",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "MiniCPM-V-4.6",
     "url": "https://huggingface.co/openbmb/MiniCPM-V-4.6",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "MiniCPM-V-4.7-35B-A3B",
     "url": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "MiniCPM-OCR-1.0",
     "url": "https://huggingface.co/openbmb/MiniCPM-OCR-1.0",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open weights on Hugging Face and code on GitHub.",
   "key_people": [
    {
     "name": "Liu Zhiyuan",
     "role": "Professor, Tsinghua University; ModelBest co-founder and Chief Scientist",
     "url": "http://nlp.csai.tsinghua.edu.cn/~lzy/",
     "note": "Associated with the Tsinghua NLP group and ModelBest; co-author of the MiniCPM paper."
    },
    {
     "name": "Maosong Sun",
     "role": "Professor, Tsinghua University",
     "url": "https://arxiv.org/abs/2404.06395",
     "note": "Senior author of the MiniCPM paper; described as a principal investigator of the community."
    },
    {
     "name": "Li Dahai",
     "role": "CEO, ModelBest",
     "url": "https://en.wikipedia.org/wiki/ModelBest",
     "note": "Co-author of the MiniCPM paper and head of the company."
    },
    {
     "name": "Shengding Hu",
     "role": "First author of MiniCPM",
     "url": "https://arxiv.org/abs/2404.06395",
     "note": "Led the MiniCPM small language model paper."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website or GitHub and Hugging Face pages.",
   "sources": [
    "https://huggingface.co/openbmb",
    "https://www.openbmb.cn"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "openbmb-modelbest",
   "mlat": 39.94523,
   "mlon": 116.45283,
   "logo": true
  },
  {
   "orig_id": "opencsg",
   "name": "OpenCSG",
   "alias": null,
   "full_name": "OpenCSG",
   "type": "startup",
   "founded": 2023,
   "country": "China",
   "city": "Beijing",
   "lat": 39.9,
   "lon": 116.41,
   "website": "https://www.opencsg.com",
   "huggingface": "https://huggingface.co/opencsg",
   "github": "https://github.com/OpenCSGs",
   "x": "https://twitter.com/OpenCsg",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Releases open Chinese datasets such as the Fineweb-Edu-Chinese series, and language models, plus a model software stack.",
   "tagline": "Initiative founded in 2023 that builds the CSGHub platform for managing models, datasets and agents, and releases open Chinese datasets and models.",
   "about": [
    "OpenCSG describes itself as an open-source and fair initiative established in 2023 to build an open community for large language models. The name stands for Converged resources, Software refined and Generative LM.",
    "Its main project is CSGHub, an Apache 2.0 platform for managing LLMs, datasets and agents with on-premise and hosted options. Other repositories cover the server, SDK, deployment charts and a lightweight local inference tool.",
    "On Hugging Face it publishes datasets and models, with Agentic-27B and Agentic-30B-A3B among its recent releases."
   ],
   "mission": "OpenCSG aims to help enterprises cut costs and raise efficiency in developing and deploying large language models through open tooling.",
   "notice": null,
   "known_for": [
    "CSGHub model and dataset platform",
    "Open Chinese datasets",
    "Agentic model releases"
   ],
   "focus": [
    "datasets",
    "language models",
    "Chinese"
   ],
   "contributions": [
    {
     "title": "OpenCSG-R1 Qwen2.5 Code model",
     "note": "A 3-billion-parameter code model published on Hugging Face in March 2025.",
     "year": "2025"
    },
    {
     "title": "Agentic-30B-A3B",
     "note": "A mixture-of-experts model published in August 2026.",
     "year": "2026"
    },
    {
     "title": "Agentic-27B",
     "note": "A 27-billion-parameter model published in September 2026.",
     "year": "2026"
    },
    {
     "title": "Agentic-27B-V2",
     "note": "A second version of the 27-billion-parameter model published in November 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "csghub",
     "url": "https://github.com/OpenCSGs/csghub",
     "note": "Open-source LLM management platform."
    },
    {
     "name": "csghub-server",
     "url": "https://github.com/OpenCSGs/csghub-server",
     "note": "Backend server for CSGHub."
    },
    {
     "name": "llm-inference",
     "url": "https://github.com/OpenCSGs/llm-inference",
     "note": "Platform for publishing LLM inference."
    },
    {
     "name": "csglite",
     "url": "https://github.com/OpenCSGs/csglite",
     "note": "Lightweight local inference platform."
    }
   ],
   "hf_highlights": [
    {
     "name": "Fineweb-Edu-Chinese-V2.1",
     "url": "https://huggingface.co/datasets/opencsg/Fineweb-Edu-Chinese-V2.1",
     "kind": "dataset",
     "note": "Chinese edu corpus."
    },
    {
     "name": "chinese-fineweb-edu",
     "url": "https://huggingface.co/datasets/opencsg/chinese-fineweb-edu",
     "kind": "dataset",
     "note": "Original Chinese edu corpus."
    },
    {
     "name": "Fineweb-Edu-Chinese-V2.2",
     "url": "https://huggingface.co/datasets/opencsg/Fineweb-Edu-Chinese-V2.2",
     "kind": "dataset",
     "note": "Chinese edu corpus."
    },
    {
     "name": "Agentic-30B-A3B",
     "url": "https://huggingface.co/opencsg/Agentic-30B-A3B",
     "kind": "model",
     "note": "MoE agentic model."
    }
   ],
   "papers": [
    {
     "title": "OpenCSG Chinese Corpus: A Series of High-quality Chinese Datasets for LLM Training",
     "url": "https://arxiv.org/abs/2501.08197",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Fineweb-Edu-Chinese-V3",
     "url": "https://huggingface.co/datasets/opencsg/Fineweb-Edu-Chinese-V3",
     "note": "Dataset"
    },
    {
     "name": "Agentic-27B-V2",
     "url": "https://huggingface.co/opencsg/Agentic-27B-V2",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Datasets and models on Hugging Face; stack on GitHub.",
   "key_people": [],
   "discord_note": null,
   "contact_note": "Contact via its website or GitHub issues.",
   "sources": [
    "https://huggingface.co/opencsg",
    "https://www.opencsg.com"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "opencsg",
   "mlat": 39.94523,
   "mlon": 116.36717,
   "logo": true
  },
  {
   "orig_id": "openeurollm",
   "name": "OpenEuroLLM",
   "alias": null,
   "full_name": "OpenEuroLLM",
   "type": "collective",
   "founded": 2025,
   "country": "Czechia",
   "city": "Prague",
   "lat": 50.0755,
   "lon": 14.4378,
   "website": "https://openeurollm.eu",
   "huggingface": "https://huggingface.co/openeurollm",
   "github": null,
   "x": "https://x.com/OpenEuroLLM",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/open-euro-llm",
   "contact_email": null,
   "summary": "EU-funded consortium of European research groups and companies building open multilingual foundation models for European languages. It publishes models, data and evaluations on Hugging Face.",
   "tagline": "EU-funded consortium of universities, companies and supercomputing centres building open, multilingual foundation models for the languages of Europe.",
   "about": [
    "OpenEuroLLM is co-funded by the European Union through the Digital Europe Programme. Its consortium of 21 organizations includes 12 universities and research organizations, six companies and three high-performance computing centres.",
    "Press and university sources describe it as coordinated by Jan Hajič of Charles University and co-led by Peter Sarlin of AMD Silo AI, starting in February 2025 for 36 months. The project publishes models, data and evaluation results on Hugging Face and GitHub.",
    "Its goals include strong multilingual models for EU official languages, wider evaluation including AI Act alignment, open datasets, open tools and recipes, and a community of developers. Recent Hugging Face releases include the Prelude9B work and numerous data-mix experiments."
   ],
   "mission": "OpenEuroLLM aims to provide transparent, performant models for European languages while preserving linguistic and cultural diversity, and to share the data, code and recipes openly.",
   "notice": null,
   "known_for": [
    "EU-funded open multilingual LLM effort",
    "21-partner European consortium",
    "Prelude9B models",
    "Open datasets and evaluations for EU languages"
   ],
   "focus": [
    "language models",
    "multilingual",
    "European languages",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Project launch",
     "note": "The consortium launched in Prague, with Charles University as coordinator; the official start was in February 2025.",
     "year": "2025"
    },
    {
     "title": "First Winter School",
     "note": "A winter school on multilinguality in LLM development and evaluation was held in Norway in February 2026.",
     "year": "2026"
    },
    {
     "title": "Open Source Sovereign Foundation Models workshop",
     "note": "Project partners presented goals and status at the OSFM workshop in Cologne in May 2026.",
     "year": "2026"
    },
    {
     "title": "Prelude models on Hugging Face",
     "note": "Prelude-long and many data-mix experiments were published in November 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "openeurollm/datamix-2b-80-20",
     "url": "https://huggingface.co/openeurollm/datamix-2b-80-20",
     "kind": "model",
     "note": "Model: 5,848 downloads, 0 likes."
    },
    {
     "name": "openeurollm/prelude-base",
     "url": "https://huggingface.co/openeurollm/prelude-base",
     "kind": "model",
     "note": "Model: 3,935 downloads, 21 likes."
    },
    {
     "name": "openeurollm/propella-annotations",
     "url": "https://huggingface.co/datasets/openeurollm/propella-annotations",
     "kind": "dataset",
     "note": "Dataset: 7,820 downloads, 22 likes."
    },
    {
     "name": "openeurollm/oellm_progress",
     "url": "https://huggingface.co/spaces/openeurollm/oellm_progress",
     "kind": "space",
     "note": "Space: 29 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "prelude-base",
     "url": "https://huggingface.co/openeurollm/prelude-base",
     "note": "Base model checkpoint from the project"
    },
    {
     "name": "propella annotations",
     "url": "https://huggingface.co/datasets/openeurollm/propella-annotations",
     "note": "Document annotations dataset"
    },
    {
     "name": "OELLM progress",
     "url": "https://huggingface.co/spaces/openeurollm/oellm_progress",
     "note": "Progress dashboard"
    }
   ],
   "openness": "The project states open-source goals and releases models, datasets and evaluation tools on Hugging Face.",
   "key_people": [
    {
     "name": "Jan Hajič",
     "role": "Project coordinator, Charles University",
     "url": "https://www.mff.cuni.cz/en/public/news/charles-university-as-the-main-coordinator-of-the-openeurollm-project",
     "note": null
    },
    {
     "name": "Peter Sarlin",
     "role": "Co-lead, AMD Silo AI",
     "url": "https://www.amd.com/en/blogs/2025/open-llms-for-transparent-ai-in-europe.html",
     "note": null
    },
    {
     "name": "Stephan Oepen",
     "role": "Partner representative, University of Oslo",
     "url": "https://openeurollm.eu",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the project website, LinkedIn or the Hugging Face org.",
   "sources": [
    "https://openeurollm.eu",
    "https://huggingface.co/openeurollm"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "openeurollm",
   "mlat": 50.09715,
   "mlon": 14.37937,
   "logo": true
  },
  {
   "orig_id": "opengpt-x",
   "name": "OpenGPT-X",
   "alias": null,
   "full_name": "OpenGPT-X",
   "type": "collective",
   "founded": 2022,
   "country": "Germany",
   "city": "Sankt Augustin",
   "lat": 50.7667,
   "lon": 7.1833,
   "website": "https://opengpt-x.de",
   "huggingface": "https://huggingface.co/openGPT-X",
   "github": "https://github.com/OpenGPTX",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@iais.fraunhofer.de",
   "summary": "German publicly funded research consortium that trained the Teuken open multilingual language models for European languages.",
   "tagline": "German publicly funded consortium that trained the open Teuken-7B models, built from scratch for all 24 official languages of the European Union.",
   "about": [
    "OpenGPT-X was a collaborative project funded by the German Federal Ministry for Economic Affairs and Climate Action from January 2022 to December 2024. It brought together partners from industry, research and media, with Fraunhofer IAIS and IIS leading and compute from the Jülich Supercomputing Centre.",
    "Its main output was Teuken-7B, a seven-billion-parameter model trained from scratch on the 24 official EU languages, with about half of its pretraining data not in English. Instruct and base versions appeared on Hugging Face from September 2024, and the project issued its public press release on 26 November 2024.",
    "The team also published a European LLM Leaderboard in July 2024 and an LLM Workbook, and Deutsche Telekom's T-Systems became the first provider to move Teuken from research to commercial use."
   ],
   "mission": "OpenGPT-X set out to develop versatile, trustworthy, multilingual and open language models made in Germany and to support their adoption in the German economy.",
   "notice": null,
   "known_for": [
    "Teuken-7B multilingual European model",
    "24 EU languages trained from scratch",
    "European LLM Leaderboard",
    "Publicly funded open model project"
   ],
   "focus": [
    "language models",
    "multilingual",
    "European languages"
   ],
   "contributions": [
    {
     "title": "Project start",
     "note": "The consortium began work in January 2022 with federal funding.",
     "year": "2022"
    },
    {
     "title": "European LLM Leaderboard",
     "note": "The team published a leaderboard for European-language model evaluation in July 2024.",
     "year": "2024"
    },
    {
     "title": "Teuken-7B research releases",
     "note": "Instruct versions v0.4 appeared on Hugging Face in September and October 2024.",
     "year": "2024"
    },
    {
     "title": "Teuken-7B public release",
     "note": "Fraunhofer announced the open release of Teuken-7B on 26 November 2024.",
     "year": "2024"
    },
    {
     "title": "Teuken-7B v0.6",
     "note": "Updated base and instruct versions published between December 2024 and July 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "lm-evaluation-harness",
     "url": "https://github.com/OpenGPTX/lm-evaluation-harness",
     "note": "Fork of the few-shot evaluation framework (13 stars)"
    },
    {
     "name": "illuminer",
     "url": "https://github.com/OpenGPTX/illuminer",
     "note": "Data quality tooling (11 stars)"
    },
    {
     "name": "Megatron-LM",
     "url": "https://github.com/OpenGPTX/Megatron-LM",
     "note": "Fork for training transformers at scale (4 stars)"
    },
    {
     "name": "AP5",
     "url": "https://github.com/OpenGPTX/AP5",
     "note": "Project repository (2 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "openGPT-X/Teuken-7B-instruct-research-v0.4",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-research-v0.4",
     "kind": "model",
     "note": "Instruct model for research use, the most downloaded OpenGPT-X model."
    },
    {
     "name": "openGPT-X/Teuken-7B-instruct-commercial-v0.4",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-commercial-v0.4",
     "kind": "model",
     "note": "Instruct model with commercial-use licence."
    },
    {
     "name": "openGPT-X/Teuken-7B-instruct-v0.6",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-v0.6",
     "kind": "model",
     "note": "Instruct model, version 0.6."
    },
    {
     "name": "openGPT-X/leaderboard_data",
     "url": "https://huggingface.co/datasets/openGPT-X/leaderboard_data",
     "kind": "dataset",
     "note": "Leaderboard data for multilingual evaluation."
    }
   ],
   "papers": [
    {
     "title": "Teuken-7B-Base & Teuken-7B-Instruct: Towards European LLMs",
     "url": "https://arxiv.org/abs/2410.03730",
     "year": "2024"
    },
    {
     "title": "Tokenizer Choice For LLM Training: Negligible or Crucial?",
     "url": "https://arxiv.org/abs/2310.08754",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "Teuken-7B-instruct-research",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-research-v0.4",
     "note": "Instruction model for 24 European languages"
    },
    {
     "name": "Teuken-7B-instruct-commercial",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-commercial-v0.4",
     "note": "Commercially usable variant"
    },
    {
     "name": "Teuken-7B-instruct-v0.6",
     "url": "https://huggingface.co/openGPT-X/Teuken-7B-instruct-v0.6",
     "note": "Updated model"
    }
   ],
   "openness": "Teuken weights are public on Hugging Face; some variants have research-only licences.",
   "key_people": [
    {
     "name": "Nicolas Flores-Herr",
     "role": "Project lead, Fraunhofer IAIS",
     "url": "https://arxiv.org/abs/2410.03730",
     "note": "Identified as OpenGPT-X project lead and co-author of the Teuken-7B paper."
    },
    {
     "name": "Georg Rehm",
     "role": "Researcher, DFKI",
     "url": "https://arxiv.org/abs/2410.03730",
     "note": "Co-author of the Teuken-7B paper and speaker at the 2024 OpenGPT-X Forum."
    },
    {
     "name": "Stefan Kesselheim",
     "role": "Researcher, Jülich Supercomputing Centre",
     "url": "https://arxiv.org/abs/2410.03730",
     "note": "Co-author of the Teuken-7B paper; Jülich supplied the training compute."
    },
    {
     "name": "Mehdi Ali",
     "role": "Lead author, Teuken-7B paper",
     "url": "https://arxiv.org/abs/2410.03730",
     "note": "First author of the Teuken-7B technical paper."
    }
   ],
   "discord_note": null,
   "contact_note": "The project's contact page names the Fraunhofer IAIS address for questions about the project.",
   "sources": [
    "https://opengpt-x.de",
    "https://huggingface.co/openGPT-X",
    "https://github.com/OpenGPTX",
    "https://arxiv.org/abs/2410.03730"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "opengpt-x",
   "logo": true
  },
  {
   "orig_id": "openhands",
   "name": "OpenHands",
   "alias": "All Hands AI",
   "full_name": "OpenHands (All Hands AI)",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "Boston",
   "lat": 42.3601,
   "lon": -71.0589,
   "website": "https://www.openhands.dev",
   "huggingface": "https://huggingface.co/OpenHands",
   "github": "https://github.com/OpenHands",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source platform and company for AI software-development agents, formerly All Hands AI.",
   "tagline": "Open-source platform for AI software-engineering agents, built by the former All Hands AI team and used by hundreds of contributors.",
   "about": [
    "OpenHands is an open-source platform for building and running AI coding agents. Its main repository, created in March 2024, is released under the MIT license and has more than 90,000 GitHub stars.",
    "The company behind it offers a GUI, a command-line tool, an SDK, a cloud service and enterprise deployments with control over models, data and access. Its site lists more than 500 open-source contributors and more than nine million downloads.",
    "The team is led by Robert Brennan as CEO, Graham Neubig as Chief Scientist and Xingyao Wang as Chief AI Officer."
   ],
   "mission": "OpenHands aims to let every team build software autonomously with open, model-agnostic agents that can run wherever they are deployed.",
   "notice": null,
   "known_for": [
    "Open-source coding agent platform",
    "SWE-bench-style agent research",
    "Model-agnostic CLI, GUI and SDK",
    "Large open contributor community"
   ],
   "focus": [
    "coding agents",
    "software engineering",
    "agents"
   ],
   "contributions": [
    {
     "title": "Launched as OpenDevin",
     "note": "The project started in March 2024 as an open alternative to Devin.",
     "year": "2024"
    },
    {
     "title": "OpenHands paper",
     "note": "The platform paper appeared on arXiv in July 2024.",
     "year": "2024"
    },
    {
     "title": "Renamed OpenHands and seed round",
     "note": "The project was renamed in late summer 2024 and All Hands AI announced a $5 million seed round on September 5, 2024.",
     "year": "2024"
    },
    {
     "title": "Series A",
     "note": "Press reports an $18.8 million Series A led by Madrona in November 2025.",
     "year": "2025"
    },
    {
     "title": "90,000 GitHub stars",
     "note": "The repository passed 90,000 stars.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "OpenHands",
     "url": "https://github.com/OpenHands/OpenHands",
     "note": "AI-driven development platform"
    }
   ],
   "hf_highlights": [
    {
     "name": "openhands-lm-32b-v0.1-ep3",
     "url": "https://huggingface.co/OpenHands/openhands-lm-32b-v0.1-ep3",
     "kind": "model",
     "note": "32B open coding model"
    }
   ],
   "papers": [
    {
     "title": "OpenHands: An Open Platform for AI Software Developers as Generalist Agents",
     "url": "https://arxiv.org/abs/2407.16741",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "OpenHands",
     "url": "https://github.com/OpenHands/OpenHands",
     "note": "AI-driven development platform"
    }
   ],
   "openness": "MIT-licensed OpenHands platform.",
   "key_people": [
    {
     "name": "Robert Brennan",
     "role": "CEO",
     "url": "https://www.openhands.dev/about",
     "note": "Leads the company."
    },
    {
     "name": "Graham Neubig",
     "role": "Chief Scientist",
     "url": "https://www.phontron.com/",
     "note": "Senior author of the OpenHands paper; Carnegie Mellon professor."
    },
    {
     "name": "Xingyao Wang",
     "role": "Chief AI Officer",
     "url": "http://xwang.dev/",
     "note": "First author of the OpenHands paper; co-founder of All Hands AI."
    },
    {
     "name": "Chuck Butkus",
     "role": "Principal Engineer",
     "url": "https://www.openhands.dev/about",
     "note": "Listed on the team page."
    },
    {
     "name": "Tim O'Farrell",
     "role": "Founding Principal Engineer",
     "url": "https://www.openhands.dev/about",
     "note": "Listed on the team page."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.openhands.dev/",
    "https://github.com/OpenHands/OpenHands",
    "https://huggingface.co/OpenHands"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "openhands",
   "logo": true
  },
  {
   "orig_id": "openllm-france",
   "name": "OpenLLM France",
   "alias": null,
   "full_name": "OpenLLM France",
   "type": "collective",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.openllm-france.fr",
   "huggingface": "https://huggingface.co/OpenLLM-France",
   "github": "https://github.com/OpenLLM-France",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "contact@openllm-france.fr",
   "summary": "French open-source community building open French-language models and datasets. It released the Lucie and Luciole models.",
   "tagline": "French community and BPI-funded consortium building open commons for French, including the Lucie-7B and Luciole model families.",
   "about": [
    "OpenLLM France grew out of a community of academic and industry stakeholders interested in open-source generative models, with a GitHub organization created in June 2023. Its current research project is funded by BPI France for two years and run by a consortium of nine official partners with 12 associate partners.",
    "Lucie-7B, published in October 2024 under Apache 2.0, was the first foundation model the project trained from scratch, with more than 30 percent French training data. The newer Luciole family covers 1B, 8B and 23B models with about 30 percent French data, with training data and code released on Hugging Face and GitHub.",
    "Public GitHub members include researchers affiliated with Linagora and CNRS."
   ],
   "mission": "The project aims to build open digital commons for the French language and share expertise for ethical, open AI applications in France.",
   "notice": null,
   "known_for": [
    "Lucie-7B French-focused foundation model",
    "Luciole open model family",
    "Open French training data and code"
   ],
   "focus": [
    "language models",
    "French",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Community founded",
     "note": "The OpenLLM France GitHub organization was created in June 2023.",
     "year": "2023"
    },
    {
     "title": "Lucie-7B",
     "note": "The first model trained from scratch by the project, released under Apache 2.0 in October 2024.",
     "year": "2024"
    },
    {
     "title": "Luciole-1B",
     "note": "A small base model released under Apache 2.0 in February 2026.",
     "year": "2026"
    },
    {
     "title": "Luciole thinking models",
     "note": "Reasoning variants including Luciole-23B-Thinking appeared in October and November 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "Manifesto",
     "url": "https://github.com/OpenLLM-France/Manifesto",
     "note": "Page de préconfiguration de la communauté OpenLLM-France (53 stars)"
    },
    {
     "name": "Lucie-Training",
     "url": "https://github.com/OpenLLM-France/Lucie-Training",
     "note": "Code for continual pretraining of LUCIE (53 stars)"
    },
    {
     "name": "Lit-Claire",
     "url": "https://github.com/OpenLLM-France/Lit-Claire",
     "note": "Continual pretraining of foundation LLM using ⚡ Lightning Fabric (37 stars)"
    },
    {
     "name": "Claire-datasets",
     "url": "https://github.com/OpenLLM-France/Claire-datasets",
     "note": "Lists of conversational datasets (21 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "OpenLLM-France/Luciole-8B-Instruct-1.1",
     "url": "https://huggingface.co/OpenLLM-France/Luciole-8B-Instruct-1.1",
     "kind": "model",
     "note": "Model: 3,432 downloads, 12 likes."
    },
    {
     "name": "OpenLLM-France/Lucie-7B-Instruct-human-data",
     "url": "https://huggingface.co/OpenLLM-France/Lucie-7B-Instruct-human-data",
     "kind": "model",
     "note": "Model: 1,324 downloads, 7 likes."
    },
    {
     "name": "OpenLLM-France/Lucie-Training-Dataset",
     "url": "https://huggingface.co/datasets/OpenLLM-France/Lucie-Training-Dataset",
     "kind": "dataset",
     "note": "Dataset: 18,594 downloads, 40 likes."
    },
    {
     "name": "OpenLLM-France/open-asr-leaderboard-luciole-audio",
     "url": "https://huggingface.co/spaces/OpenLLM-France/open-asr-leaderboard-luciole-audio",
     "kind": "space",
     "note": "Space: 0 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "The Lucie-7B LLM and the Lucie Training Dataset: Open resources for multilingual language generation",
     "url": "https://arxiv.org/abs/2503.12294",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Lucie-7B",
     "url": "https://huggingface.co/OpenLLM-France/Lucie-7B",
     "note": "Open multilingual model with French focus"
    },
    {
     "name": "Luciole-23B-Instruct",
     "url": "https://huggingface.co/OpenLLM-France/Luciole-23B-Instruct-1.1",
     "note": "Instruction model"
    },
    {
     "name": "Lucie Training Dataset",
     "url": "https://huggingface.co/datasets/OpenLLM-France/Lucie-Training-Dataset",
     "note": "Open training dataset"
    }
   ],
   "openness": "Models, datasets and training code are public; some code repositories use GPL or AGPL licences.",
   "key_people": [
    {
     "name": "Jérôme Louradour",
     "role": "Public org member",
     "url": "https://github.com/Jeronymous",
     "note": null
    },
    {
     "name": "Julie Hunter",
     "role": "Public org member",
     "url": "https://github.com/jhunter19",
     "note": null
    },
    {
     "name": "Christophe Cerisara",
     "role": "Public org member",
     "url": "https://github.com/cerisara",
     "note": null
    },
    {
     "name": "Michel-Marie Maudet",
     "role": "Linagora, public org member",
     "url": "http://www.linagora.com/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on the website.",
   "sources": [
    "https://www.openllm-france.fr",
    "https://huggingface.co/OpenLLM-France",
    "https://github.com/OpenLLM-France",
    "https://arxiv.org/abs/2503.12294"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "openllm-france",
   "mlat": 48.91872,
   "mlon": 2.22718,
   "logo": true
  },
  {
   "orig_id": "openllm-ro",
   "name": "OpenLLM-Ro",
   "alias": null,
   "full_name": "OpenLLM-Ro",
   "type": "collective",
   "founded": 2024,
   "country": "Romania",
   "city": "Bucharest",
   "lat": 44.4268,
   "lon": 26.1025,
   "website": "http://openllm.ro/",
   "huggingface": "https://huggingface.co/OpenLLM-Ro",
   "github": "https://github.com/OpenLLM-Ro",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Community of Romanian researchers building open Romanian language models and datasets.",
   "tagline": "Community of Romanian researchers releasing open Romanian language models, datasets, benchmarks and training code.",
   "about": [
    "OpenLLM-Ro is a community building open large language models for Romanian. Its GitHub organization was created in April 2024, and its first Romanian models, fine-tuned from Llama 2, appeared on 14 May 2024.",
    "It has since released models based on Mistral, Gemma and Llama 3, instruction-tuned and human-aligned variants, training code, evaluation code and two LLM-as-a-judge benchmarks, RoMTBench and RoCulturaBench. In 2025 and 2026 it added pretraining data with a quality classifier, vision-language data and a quantized RoLlama-3.2-1B.",
    "The group has published preprints on its models, data quality and vision-language work, and held a workshop on LLMs for Romanian in September 2024."
   ],
   "mission": "OpenLLM-Ro aims to build and share open language models and resources for Romanian.",
   "notice": null,
   "known_for": [
    "RoLlama, RoMistral and RoGemma Romanian models",
    "RoMTBench and RoCulturaBench",
    "Open Romanian training and evaluation code"
   ],
   "focus": [
    "language models",
    "Romanian",
    "datasets"
   ],
   "contributions": [
    {
     "title": "First Romanian models",
     "note": "RoLlama2 models were released on 14 May 2024.",
     "year": "2024"
    },
    {
     "title": "RoMistral, RoGemma and RoLlama3",
     "note": "Additional model families were released between May and June 2024.",
     "year": "2024"
    },
    {
     "title": "Training code and benchmarks",
     "note": "Training code (August 2024) and RoMTBench with RoCulturaBench (August 2024) were published.",
     "year": "2024"
    },
    {
     "title": "Pretraining data and quality classifier",
     "note": "Augmented Romanian pretraining data was released in November 2025.",
     "year": "2025"
    },
    {
     "title": "Vision-language resources",
     "note": "Code, training and evaluation data for Romanian vision-language models were released in June 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "llama-recipes",
     "url": "https://github.com/OpenLLM-Ro/llama-recipes",
     "note": "Recipes used for training Romanian models (23 stars)"
    },
    {
     "name": "LLaMA-Factory",
     "url": "https://github.com/OpenLLM-Ro/LLaMA-Factory",
     "note": "Fork with code and data used for training Romanian models (11 stars)"
    },
    {
     "name": "FastChat",
     "url": "https://github.com/OpenLLM-Ro/FastChat",
     "note": "Fork for evaluating Romanian LLMs with LLM-as-a-judge (2 stars)"
    },
    {
     "name": "lm-evaluation-harness",
     "url": "https://github.com/OpenLLM-Ro/lm-evaluation-harness",
     "note": "Fork with code and data for evaluating Romanian LLMs (1 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "OpenLLM-Ro/RoLlama3.1-8b-Instruct",
     "url": "https://huggingface.co/OpenLLM-Ro/RoLlama3.1-8b-Instruct",
     "kind": "model",
     "note": "Romanian instruct model based on Llama 3.1."
    },
    {
     "name": "OpenLLM-Ro/RoGemma2-9b-Instruct",
     "url": "https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct",
     "kind": "model",
     "note": "Romanian instruct model based on Gemma 2."
    },
    {
     "name": "OpenLLM-Ro/ro_mmlu",
     "url": "https://huggingface.co/datasets/OpenLLM-Ro/ro_mmlu",
     "kind": "dataset",
     "note": "Romanian MMLU evaluation set."
    },
    {
     "name": "OpenLLM-Ro/fineweb2-ro-bert",
     "url": "https://huggingface.co/datasets/OpenLLM-Ro/fineweb2-ro-bert",
     "kind": "dataset",
     "note": "Romanian pretraining data with quality annotations."
    }
   ],
   "papers": [
    {
     "title": "\"Vorbeşti Româneşte?\" A Recipe to Train Powerful Romanian LLMs with English Instructions",
     "url": "https://arxiv.org/abs/2406.18266",
     "year": "2024"
    },
    {
     "title": "\"Înţelegi Româneşte?\" A Recipe for Romanian Vision-Language Models",
     "url": "https://arxiv.org/abs/2605.31401",
     "year": "2026"
    },
    {
     "title": "Improving Romanian LLM Pretraining Data using Diversity and Quality Filtering",
     "url": "https://arxiv.org/abs/2511.01090",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "RoLlama3-8b-Instruct",
     "url": "https://huggingface.co/OpenLLM-Ro/RoLlama3-8b-Instruct-2024-06-28",
     "note": "Romanian instruction model"
    },
    {
     "name": "ro_sft_alpaca",
     "url": "https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_alpaca",
     "note": "Romanian instruction data"
    },
    {
     "name": "ro_wiki",
     "url": "https://huggingface.co/datasets/OpenLLM-Ro/ro_wiki",
     "note": "Romanian Wikipedia dataset"
    }
   ],
   "openness": "Models, datasets and evaluation code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Mihai Masala",
     "role": "Lead author, \"Vorbeşti Româneşte?\"",
     "url": "https://arxiv.org/abs/2406.18266",
     "note": "First author of the RoLlama recipe paper (EMNLP 2024 Findings)."
    },
    {
     "name": "Traian Rebedea",
     "role": "Researcher; co-author",
     "url": "https://arxiv.org/abs/2511.01090",
     "note": "Senior author of the Romanian pretraining data quality paper."
    },
    {
     "name": "Mihai Dascalu",
     "role": "Researcher; co-author",
     "url": "https://arxiv.org/abs/2406.18266",
     "note": "Co-author of the founding paper."
    },
    {
     "name": "Marius Leordeanu",
     "role": "Researcher; co-author",
     "url": "https://arxiv.org/abs/2406.18266",
     "note": "Co-author of the founding paper."
    }
   ],
   "discord_note": null,
   "contact_note": "The site invites people to join the community and links to its GitHub and Hugging Face pages; no email is printed.",
   "sources": [
    "http://openllm.ro/",
    "https://huggingface.co/OpenLLM-Ro",
    "https://github.com/OpenLLM-Ro",
    "https://arxiv.org/abs/2406.18266"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "openllm-ro",
   "logo": true
  },
  {
   "orig_id": "openmoss",
   "name": "OpenMOSS",
   "alias": null,
   "full_name": "OpenMOSS",
   "type": "academic",
   "founded": 2023,
   "country": "China",
   "city": "Shanghai",
   "lat": 31.23,
   "lon": 121.47,
   "website": "http://openmoss.sii.edu.cn/",
   "huggingface": "https://huggingface.co/OpenMOSS-Team",
   "github": "https://github.com/OpenMOSS",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "LLM research group at the Shanghai Innovation Institution working with Fudan University, led by Prof. Xipeng Qiu.",
   "tagline": "Fudan-linked research group at the Shanghai Innovation Institute behind the MOSS models, MOSS-TTS speech and image releases and interpretability tools.",
   "about": [
    "OpenMOSS is a research group under the Shanghai Innovation Institute, working closely with Fudan University and MOSS-AI. Its GitHub profile lists Xipeng Qiu of Fudan as an associated academic, and its site describes work on open-source large models and AGI research.",
    "The MOSS repository was created in April 2023 and has more than 12,000 stars under Apache 2.0. More recent releases include MOSS-TTS in February 2026, MOSS-TTS-Nano in April 2026, MOSS-Transcribe-Diarize in May 2026 and MOSS-Image in October 2026.",
    "The team also publishes world-action-model work under the EasyWAM name and maintains the Awesome-WAM list."
   ],
   "mission": "OpenMOSS conducts open research on large models and general artificial intelligence and releases its code and models openly.",
   "notice": null,
   "known_for": [
    "MOSS conversational language model",
    "MOSS-TTS speech models",
    "Open-source large model research at Fudan"
   ],
   "focus": [
    "language models",
    "speech",
    "image generation"
   ],
   "contributions": [
    {
     "title": "MOSS repository",
     "note": "The MOSS model repository was created in April 2023 under Apache 2.0.",
     "year": "2023"
    },
    {
     "title": "MOSS-TTS",
     "note": "A speech generation model published in February 2026.",
     "year": "2026"
    },
    {
     "title": "MOSS-Transcribe-Diarize",
     "note": "A transcription and diarization model published in May 2026.",
     "year": "2026"
    },
    {
     "title": "EasyWAM models",
     "note": "World-action-model checkpoints published on Hugging Face in September 2026.",
     "year": "2026"
    },
    {
     "title": "MOSS-Image",
     "note": "Base and SFT image models released in October 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "MOSS",
     "url": "https://github.com/OpenMOSS/MOSS",
     "note": "Tool-augmented conversational model."
    },
    {
     "name": "MOSS-TTS-Nano",
     "url": "https://github.com/OpenMOSS/MOSS-TTS-Nano",
     "note": "Small multilingual TTS for CPU."
    },
    {
     "name": "MOSS-TTS",
     "url": "https://github.com/OpenMOSS/MOSS-TTS",
     "note": "Speech generation model family."
    },
    {
     "name": "MOSS-Image",
     "url": "https://github.com/OpenMOSS/MOSS-Image",
     "note": "5B DiT image model."
    }
   ],
   "hf_highlights": [
    {
     "name": "MOSS-Audio-Tokenizer-Nano",
     "url": "https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano",
     "kind": "model",
     "note": "Audio tokenizer."
    },
    {
     "name": "MOSS-TTS-Local-Transformer",
     "url": "https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Local-Transformer",
     "kind": "model",
     "note": "TTS model."
    },
    {
     "name": "MOSS-Transcribe-Diarize",
     "url": "https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize",
     "kind": "model",
     "note": "Transcription with diarization."
    },
    {
     "name": "OmniAction",
     "url": "https://huggingface.co/datasets/OpenMOSS-Team/OmniAction",
     "kind": "dataset",
     "note": "Action dataset."
    }
   ],
   "papers": [
    {
     "title": "AnyGPT: Unified Multimodal LLM with Discrete Sequence Modeling",
     "url": "https://arxiv.org/abs/2402.12226",
     "year": "2024"
    },
    {
     "title": "SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities",
     "url": "https://arxiv.org/abs/2305.11000",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "MOSS-Transcribe-Diarize",
     "url": "https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "OmniVAE",
     "url": "https://huggingface.co/OpenMOSS-Team/OmniVAE",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Models on Hugging Face, code on GitHub.",
   "key_people": [
    {
     "name": "Xipeng Qiu",
     "role": "Professor, Fudan University",
     "url": "https://xpqiu.github.io/",
     "note": null
    },
    {
     "name": "Peng Li",
     "role": "Public org member",
     "url": "https://artpli.github.io",
     "note": null
    },
    {
     "name": "Zhengfu He",
     "role": "Public org member, SII",
     "url": "https://github.com/Hzfinfdu",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the group's website or GitHub issues.",
   "sources": [
    "https://huggingface.co/OpenMOSS-Team",
    "http://openmoss.sii.edu.cn/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "openmoss",
   "mlat": 31.23,
   "mlon": 121.52847,
   "logo": true
  },
  {
   "orig_id": "open-thoughts",
   "name": "OpenThoughts",
   "alias": null,
   "full_name": "OpenThoughts",
   "type": "academic",
   "founded": 2025,
   "country": "United States",
   "city": "Mountain View",
   "lat": 37.3861,
   "lon": -122.0839,
   "website": "https://www.openthoughts.ai",
   "huggingface": "https://huggingface.co/open-thoughts",
   "github": "https://github.com/open-thoughts",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Open research collective curating fully open datasets and recipes for reasoning models, led with Bespoke Labs and academic partners; releases OpenThoughts datasets and OpenThinker models.",
   "tagline": "A research collective curating fully open reasoning datasets and recipes, from the OpenThoughts3 data recipe to the OpenThinker models.",
   "about": [
    "OpenThoughts is a collaboration of researchers and engineers from universities including Stanford, UC Berkeley, UT Austin, NYU, the University of Washington and TUM, together with LAION. It launched publicly in January 2025 with the aim of producing open post-training data for reasoning models.",
    "The project publishes datasets, training code and models together. Its OpenThoughts3 recipe, described in the paper \"OpenThoughts: Data Recipes for Reasoning Models\", was derived from more than 1,000 controlled experiments and produced the 1.2 million example OpenThoughts3 dataset and the OpenThinker3-7B model.",
    "Compute and support come from academic centers such as the Jülich Supercomputing Centre, TACC and TU Dresden, as well as industry partners. From late 2025 the group extended its work to agentic tasks under the OpenThoughts-Agent project."
   ],
   "mission": "OpenThoughts aims to make the data side of reasoning models as open as the weights, so that the community can study which data choices matter. It treats systematic, reproducible ablation of data pipelines as its main method.",
   "notice": null,
   "known_for": [
    "OpenThoughts3 reasoning dataset",
    "OpenThinker open reasoning models",
    "Large-scale data-recipe ablations",
    "OpenThoughts-Agent project"
   ],
   "focus": [
    "reasoning data",
    "datasets",
    "post-training",
    "agents"
   ],
   "contributions": [
    {
     "title": "Project launch",
     "note": "The OpenThoughts project was announced with the goal of open reasoning data.",
     "year": "2025"
    },
    {
     "title": "OpenThinker-32B",
     "note": "Released an open-data 32B reasoning model, described by the project as scaling up open reasoning.",
     "year": "2025"
    },
    {
     "title": "OpenThinker2",
     "note": "Published a follow-up model family reported to outperform DeepSeek-R1-Distill-32B on its benchmarks.",
     "year": "2025"
    },
    {
     "title": "OpenThoughts3",
     "note": "Released the 1.2M example dataset and a recipe based on over 1,000 experiments, with OpenThinker3-7B.",
     "year": "2025"
    },
    {
     "title": "OpenThoughts-Agent",
     "note": "Launched a project on data recipes and benchmarks for terminal and agentic tasks, followed by the OpenThoughts-TBLite benchmark in 2026.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "open-thoughts",
     "url": "https://github.com/open-thoughts/open-thoughts",
     "note": "Fully open data curation for reasoning models"
    },
    {
     "name": "OpenThoughts-Agent",
     "url": "https://github.com/open-thoughts/OpenThoughts-Agent",
     "note": "Data recipes and infrastructure for training agents"
    },
    {
     "name": "OpenThoughts-TBLite",
     "url": "https://github.com/open-thoughts/OpenThoughts-TBLite",
     "note": "Benchmark for terminal agents"
    }
   ],
   "hf_highlights": [
    {
     "name": "OpenThoughts-114k",
     "url": "https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k",
     "kind": "dataset",
     "note": "Reasoning dataset"
    },
    {
     "name": "OpenThoughts3-1.2M",
     "url": "https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M",
     "kind": "dataset",
     "note": "1.2M-example reasoning dataset"
    },
    {
     "name": "OpenThinker3-7B",
     "url": "https://huggingface.co/open-thoughts/OpenThinker3-7B",
     "kind": "model",
     "note": "7B reasoning model"
    },
    {
     "name": "OpenThinker-Agent-v1",
     "url": "https://huggingface.co/open-thoughts/OpenThinker-Agent-v1",
     "kind": "model",
     "note": "Terminal-agent model"
    }
   ],
   "papers": [
    {
     "title": "OpenThoughts: Data Recipes for Reasoning Models",
     "url": "https://arxiv.org/abs/2506.04178",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "open-thoughts",
     "url": "https://github.com/open-thoughts/open-thoughts",
     "note": "Fully open data curation for reasoning models"
    },
    {
     "name": "OpenThoughts-Agent",
     "url": "https://github.com/open-thoughts/OpenThoughts-Agent",
     "note": "Data recipes and infrastructure for training agents"
    },
    {
     "name": "OpenThoughts-TBLite",
     "url": "https://github.com/open-thoughts/OpenThoughts-TBLite",
     "note": "Benchmark for terminal agents"
    }
   ],
   "openness": "Apache-2.0 code, with datasets and OpenThinker model weights released on Hugging Face.",
   "key_people": [
    {
     "name": "Etash Guha",
     "role": "Lead author, OpenThoughts",
     "url": "https://arxiv.org/abs/2506.04178",
     "note": "First author of the OpenThoughts data recipes paper."
    },
    {
     "name": "Ludwig Schmidt",
     "role": "Assistant professor, Stanford; LAION member",
     "url": "https://people.csail.mit.edu/ludwigs/",
     "note": "Last author of the OpenThoughts paper."
    },
    {
     "name": "Alexandros G. Dimakis",
     "role": "Professor, UC Berkeley; Bespoke Labs co-founder",
     "url": "https://people.eecs.berkeley.edu/~alexdimakis/",
     "note": "Co-author of OpenThoughts and OpenThoughts-Agent."
    },
    {
     "name": "Ryan Marten",
     "role": "Co-author",
     "url": "https://arxiv.org/abs/2506.04178",
     "note": "Second author of the OpenThoughts paper."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://www.openthoughts.ai",
    "https://github.com/open-thoughts",
    "https://huggingface.co/open-thoughts",
    "https://bespokelabs.ai/"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "openthoughts",
   "mlat": 37.40775,
   "mlon": -122.1311,
   "logo": true
  },
  {
   "orig_id": "oumi",
   "name": "Oumi",
   "alias": null,
   "full_name": "Oumi",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "Seattle",
   "lat": 47.6062,
   "lon": -122.3321,
   "website": "https://oumi.ai",
   "huggingface": "https://huggingface.co/oumi-ai",
   "github": "https://github.com/oumi-ai",
   "x": "https://x.com/Oumi_ai",
   "discord": "https://discord.com/invite/oumi",
   "linkedin": null,
   "contact_email": null,
   "summary": "Public-benefit company behind an unconditionally open-source platform for training, evaluating and deploying foundation models.",
   "tagline": "A public benefit company building an unconditionally open platform for training, evaluating and deploying foundation models, with an open-source library at its core.",
   "about": [
    "Oumi emerged from stealth in January 2025 with a $10 million seed round and an open-source platform covering the model lifecycle, from data preparation and training to evaluation and deployment. It was built with researchers from more than a dozen universities.",
    "The company is led by co-founder and CEO Manos Koukoumidis, a former Google Cloud engineering manager, with co-founder Oussama Elachqar and other founders who say they worked on PaLM at Google. Oumi PBC publishes its core library as open source and has released models such as the HallOumi hallucination detector.",
    "Alongside the open library, Oumi offers a commercial platform with an agent that automates evaluation, data curation and training for teams building specialized models."
   ],
   "mission": "Oumi wants AI development to be collective and reproducible, with open code, data and models rather than closed stacks. It describes its aim as giving AI a shared, Linux-like open foundation.",
   "notice": null,
   "known_for": [
    "Unconditionally open AI platform",
    "HallOumi hallucination detection model",
    "Open library for training and evaluation",
    "Academic-industry collaboration"
   ],
   "focus": [
    "training platform",
    "evaluation",
    "hallucination detection",
    "fine-tuning"
   ],
   "contributions": [
    {
     "title": "Emerged from stealth",
     "note": "Oumi launched as a public benefit corporation on January 29, 2025 with a $10 million seed round.",
     "year": "2025"
    },
    {
     "title": "Public launch of the open platform",
     "note": "The open-source platform covers data preparation, training, evaluation and deployment.",
     "year": "2025"
    },
    {
     "title": "HallOumi",
     "note": "An open claim-verification model for detecting hallucinations, released in April 2025.",
     "year": "2025"
    },
    {
     "title": "Oumi OSS v0.8",
     "note": "Release in May 2026 accompanied by a case study of a 3B text-to-SQL model.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "oumi",
     "url": "https://github.com/oumi-ai/oumi",
     "note": "Fine-tune, evaluate and deploy open models"
    },
    {
     "name": "halloumi-demo",
     "url": "https://github.com/oumi-ai/halloumi-demo",
     "note": "HallOumi claim verification demo"
    }
   ],
   "hf_highlights": [
    {
     "name": "HallOumi-8B",
     "url": "https://huggingface.co/oumi-ai/HallOumi-8B",
     "kind": "model",
     "note": "Claim verification model"
    },
    {
     "name": "MetaMathQA-R1",
     "url": "https://huggingface.co/datasets/oumi-ai/MetaMathQA-R1",
     "kind": "dataset",
     "note": "Math reasoning dataset"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "oumi",
     "url": "https://github.com/oumi-ai/oumi",
     "note": "Fine-tune, evaluate and deploy open models"
    },
    {
     "name": "halloumi-demo",
     "url": "https://github.com/oumi-ai/halloumi-demo",
     "note": "HallOumi claim verification demo"
    }
   ],
   "openness": "Apache-2.0 platform; open HallOumi models and datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Manos Koukoumidis",
     "role": "Co-founder and CEO",
     "url": "https://yespress.io/emmanouil-koukoumidis",
     "note": null
    },
    {
     "name": "Oussama Elachqar",
     "role": "Co-founder",
     "url": "https://oumi.ai",
     "note": null
    },
    {
     "name": "Matthew Persons",
     "role": "Co-founder",
     "url": "https://oumi.ai",
     "note": null
    },
    {
     "name": "Jeremiah Greer",
     "role": "Co-founder",
     "url": "https://oumi.ai",
     "note": null
    },
    {
     "name": "William Zeng",
     "role": "Co-founder",
     "url": "https://oumi.ai",
     "note": null
    },
    {
     "name": "Kostas Aisopos",
     "role": "Co-founder",
     "url": "https://oumi.ai",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the website.",
   "contact_note": "Via website.",
   "sources": [
    "https://oumi.ai/",
    "https://github.com/oumi-ai/oumi",
    "https://huggingface.co/oumi-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "oumi",
   "mlat": 47.64156,
   "mlon": -122.3321,
   "logo": true
  },
  {
   "orig_id": "part-ai",
   "name": "Part AI",
   "alias": null,
   "full_name": "Part AI",
   "type": "startup",
   "founded": null,
   "country": "Iran",
   "city": "Tehran",
   "lat": 35.69,
   "lon": 51.39,
   "website": "https://partdp.ai/en",
   "huggingface": "https://huggingface.co/PartAI",
   "github": "https://github.com/partdpai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Iranian company publishing Persian Dorna and Tooka models.",
   "tagline": "A Tehran-based developer of Persian-language models, known for the Dorna decoder LLMs and the TookaBERT and Tooka-SBERT encoders.",
   "about": [
    "Part AI publishes open Persian-language models on Hugging Face. Its Dorna family consists of decoder-only models trained or fine-tuned on Persian data, including Dorna2 variants built on Llama 3.1.",
    "Its encoder line includes TookaBERT in base and large sizes, described in an arXiv paper as a step forward for Persian language understanding, and the Tooka-SBERT sentence embedding models in small and large sizes.",
    "The models serve Persian NLP developers and researchers who need open baselines for a language that is underrepresented in multilingual models."
   ],
   "mission": "Part AI works to strengthen Persian language technology by releasing models and benchmarks that others can build on.",
   "notice": null,
   "known_for": [
    "Dorna Persian LLMs",
    "TookaBERT Persian encoders",
    "Tooka-SBERT embeddings",
    "Open Persian NLP models"
   ],
   "focus": [
    "Persian",
    "language models",
    "embeddings"
   ],
   "contributions": [
    {
     "title": "TookaBERT",
     "note": "Published a family of Persian encoder models in base and large sizes, with an accompanying paper.",
     "year": "2024"
    },
    {
     "title": "Dorna",
     "note": "Released Persian-tuned decoder models starting with Dorna-Llama3-8B-Instruct.",
     "year": "2024"
    },
    {
     "title": "Dorna2",
     "note": "Followed with Dorna2 models built on Llama 3.1.",
     "year": "2024"
    },
    {
     "title": "Tooka-SBERT",
     "note": "Released small and large Persian sentence embedding models.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "PartNLP",
     "url": "https://github.com/partdpai/PartNLP",
     "note": "Persian NLP package."
    },
    {
     "name": "persian_subwords",
     "url": "https://github.com/partdpai/persian_subwords",
     "note": "Persian OCR with subwords."
    },
    {
     "name": "Part-2018-Deep-Learning-School",
     "url": "https://github.com/partdpai/Part-2018-Deep-Learning-School",
     "note": "Course materials."
    }
   ],
   "hf_highlights": [
    {
     "name": "PartAI/TookaBERT-Base",
     "url": "https://huggingface.co/PartAI/TookaBERT-Base",
     "kind": "model",
     "note": "Persian BERT."
    },
    {
     "name": "PartAI/TookaBERT-Large",
     "url": "https://huggingface.co/PartAI/TookaBERT-Large",
     "kind": "model",
     "note": "Large."
    },
    {
     "name": "PartAI/Tooka-SBERT-V2-Large",
     "url": "https://huggingface.co/PartAI/Tooka-SBERT-V2-Large",
     "kind": "model",
     "note": "Embeddings."
    },
    {
     "name": "PartAI/PSRB",
     "url": "https://huggingface.co/datasets/PartAI/PSRB",
     "kind": "dataset",
     "note": "Persian benchmark."
    }
   ],
   "papers": [
    {
     "title": "TookaBERT: A Step Forward for Persian NLU",
     "url": "https://arxiv.org/abs/2407.16382",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Dorna-Llama3-8B-Instruct",
     "url": "https://huggingface.co/PartAI/Dorna-Llama3-8B-Instruct",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Tooka-SBERT",
     "url": "https://huggingface.co/PartAI/Tooka-SBERT",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [
    {
     "name": "MohammadAli Sadraei Javaheri",
     "role": "Lead author, TookaBERT",
     "url": "https://arxiv.org/abs/2407.16382",
     "note": "First author of the TookaBERT paper."
    },
    {
     "name": "Tohid Abedini",
     "role": "Co-author, TookaBERT",
     "url": "https://arxiv.org/abs/2407.16382",
     "note": "Researcher at Part AI."
    },
    {
     "name": "Amirmohammad Salehoof",
     "role": "Co-author, TookaBERT",
     "url": "https://arxiv.org/abs/2407.16382",
     "note": "Researcher at Part AI."
    }
   ],
   "discord_note": null,
   "contact_note": "partdp.ai has a contact page.",
   "sources": [
    "https://huggingface.co/PartAI",
    "https://partdp.ai/en"
   ],
   "confidence": "medium",
   "region": "Middle East",
   "id": "part-ai",
   "mlat": 35.72536,
   "mlon": 51.39,
   "logo": true
  },
  {
   "orig_id": "photoroom",
   "name": "Photoroom",
   "alias": null,
   "full_name": "Photoroom",
   "type": "startup",
   "founded": 2019,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.photoroom.com",
   "huggingface": "https://huggingface.co/Photoroom",
   "github": "https://github.com/photoroom",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Paris-based image-editing company that also trains and releases its own text-to-image models (PRX) with open weights, plus open data tooling.",
   "tagline": "The Paris image-editing company that trains its own text-to-image models and releases them as open source under Apache 2.0.",
   "about": [
    "Photoroom was co-founded in 2019 by Matt, a former GoPro product manager, and engineer Eliot Andres, after frustration with slow tools for cutting out objects from photos. It builds an image-editing app and an API used for e-commerce visuals.",
    "In 2025 it began publishing its own text-to-image models, called PRX, together with an account of the architecture, training and post-training process. The 1.3 billion parameter model was trained from scratch and is available through the Diffusers library.",
    "A later pixel-space variant, PRX Pixel, uses raw images without a VAE and also carries an Apache 2.0 licence."
   ],
   "mission": "Photoroom wants to show the full process of training production-grade image models and to share it so others can build on it.",
   "notice": null,
   "known_for": [
    "PRX open text-to-image models",
    "Open training process write-ups",
    "AI photo editing for commerce"
   ],
   "focus": [
    "image generation",
    "diffusion",
    "computer vision"
   ],
   "contributions": [
    {
     "title": "Photoroom co-founded",
     "note": "Matt and Eliot Andres founded the company in Paris.",
     "year": "2019"
    },
    {
     "title": "PRX open-sourced",
     "note": "Photoroom released its 1.3B parameter text-to-image model and training process with an Apache 2.0 licence.",
     "year": "2025"
    },
    {
     "title": "PRX in Diffusers",
     "note": "The model is supported through a PRX pipeline in Hugging Face Diffusers.",
     "year": "2025"
    },
    {
     "title": "PRX Pixel published",
     "note": "A pixel-space image model without VAE was described on the Photoroom blog.",
     "year": "2026"
    },
    {
     "title": "PRX training series",
     "note": "Photoroom documented architecture, training and post-training in a multi-part blog series.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "datago",
     "url": "https://github.com/Photoroom/datago",
     "note": "A natively parallel dataloader for Python, written in Rust. Serving data at GB/s speeds, while covering aspect (128 stars)"
    },
    {
     "name": "fast-foreground-estimation",
     "url": "https://github.com/Photoroom/fast-foreground-estimation",
     "note": "Official repository for the paper Approximate Fast Foreground Colour Estimation. ICIP 2021. (93 stars)"
    },
    {
     "name": "PRX",
     "url": "https://github.com/Photoroom/PRX",
     "note": "Open-source repository. (92 stars)"
    },
    {
     "name": "dataroom",
     "url": "https://github.com/Photoroom/dataroom",
     "note": "Framework based on a vector dabase to store, manage and curate large image datasets (85 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "Photoroom/prx-1024-t2i-beta",
     "url": "https://huggingface.co/Photoroom/prx-1024-t2i-beta",
     "kind": "model",
     "note": "Model with 86 likes on the Hub."
    },
    {
     "name": "Photoroom/prxpixel-t2i",
     "url": "https://huggingface.co/Photoroom/prxpixel-t2i",
     "kind": "model",
     "note": "Model with 52 likes on the Hub."
    },
    {
     "name": "Photoroom/midjourney-v6-recap",
     "url": "https://huggingface.co/datasets/Photoroom/midjourney-v6-recap",
     "kind": "dataset",
     "note": "Dataset with 22 likes on the Hub."
    },
    {
     "name": "Photoroom/PRX-Pixel",
     "url": "https://huggingface.co/spaces/Photoroom/PRX-Pixel",
     "kind": "space",
     "note": "Space with 36 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "PRX text-to-image models",
     "url": "https://huggingface.co/Photoroom/prx-1024-t2i-beta",
     "note": "Open-weight text-to-image model released on the Hub under Apache-2.0"
    },
    {
     "name": "PRX code",
     "url": "https://github.com/photoroom/PRX",
     "note": "Training and inference code for the PRX models"
    },
    {
     "name": "datago",
     "url": "https://github.com/photoroom/datago",
     "note": "Rust-based parallel dataloader for large image datasets"
    },
    {
     "name": "midjourney-v6-recap",
     "url": "https://huggingface.co/datasets/Photoroom/midjourney-v6-recap",
     "note": "Recaptioned image dataset"
    }
   ],
   "openness": "Commercial product company with an open research arm: PRX model weights are published on Hugging Face (Apache-2.0 on the checked model card) and several tooling repositories are public on GitHub.",
   "key_people": [
    {
     "name": "Eliot Andres",
     "role": "Co-founder",
     "url": "https://www.photoroom.com/about",
     "note": "Engineer who co-founded the company in 2019."
    },
    {
     "name": "Jon Almazán",
     "role": "Lead engineer",
     "url": "https://huggingface.co/blog/Photoroom/prx-open-source-t2i-model",
     "note": "Author of the PRX release post."
    },
    {
     "name": "David Bertoin",
     "role": "Researcher",
     "url": "https://huggingface.co/blog/Photoroom/prx-open-source-t2i-model",
     "note": "Author of the PRX release post."
    },
    {
     "name": "Roman Frigg",
     "role": "Researcher",
     "url": "https://huggingface.co/blog/Photoroom/prx-open-source-t2i-model",
     "note": "Author of the PRX release post."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the company website.",
   "sources": [
    "https://www.photoroom.com",
    "https://huggingface.co/Photoroom",
    "https://github.com/photoroom"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "photoroom",
   "mlat": 48.88481,
   "mlon": 2.20151,
   "logo": true
  },
  {
   "orig_id": "physical-intelligence",
   "name": "Physical Intelligence",
   "alias": null,
   "full_name": "Physical Intelligence",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.pi.website",
   "huggingface": "https://huggingface.co/physical-intelligence",
   "github": "https://github.com/Physical-Intelligence",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Robotics foundation-model company that open-sources its pi0 family of vision-language-action models and training code via the openpi repository.",
   "tagline": "A robotics foundation-model company whose pi0 family of vision-language-action models and openpi code give robot learning a shared open baseline.",
   "about": [
    "Physical Intelligence builds general-purpose foundation models for robots rather than robot hardware. The company was co-founded by Karol Hausman, Sergey Levine, Chelsea Finn and others, and draws on research from Google, UC Berkeley and Stanford.",
    "Its pi0 model combines a vision-language backbone with a flow-matching action expert to generate robot actions. The company released pi0 weights and training and inference code through the openpi repository in early 2025.",
    "Researchers and robotics developers use openpi to fine-tune policies on their own robots, which makes the project one of the main open entry points to robot foundation models."
   ],
   "mission": "The company aims to build a general-purpose model that can control many kinds of robots across many tasks. It shares part of its research openly to help the field progress.",
   "notice": null,
   "known_for": [
    "pi0 vision-language-action model",
    "openpi open-source release",
    "General-purpose robot control",
    "Flow-matching action experts"
   ],
   "focus": [
    "robotics",
    "vision-language-action models",
    "foundation models"
   ],
   "contributions": [
    {
     "title": "pi0 announced",
     "note": "The company presented its first vision-language-action flow model.",
     "year": "2024"
    },
    {
     "title": "Series A of $400 million",
     "note": "Reported by secondary sources as raised in November 2024.",
     "year": "2024"
    },
    {
     "title": "openpi release",
     "note": "pi0 and training code released on February 4, 2025.",
     "year": "2025"
    },
    {
     "title": "pi0.5 and PyTorch support",
     "note": "The openpi README lists pi0.5 checkpoints and PyTorch support among its September 2025 additions.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "openpi",
     "url": "https://github.com/Physical-Intelligence/openpi",
     "note": "Open-source pi0 / pi0.5 models and code"
    },
    {
     "name": "real-time-chunking-kinetix",
     "url": "https://github.com/Physical-Intelligence/real-time-chunking-kinetix",
     "note": "Real-time action chunking experiments"
    },
    {
     "name": "aloha",
     "url": "https://github.com/Physical-Intelligence/aloha",
     "note": "ALOHA robot setup"
    }
   ],
   "hf_highlights": [
    {
     "name": "fast",
     "url": "https://huggingface.co/physical-intelligence/fast",
     "kind": "model",
     "note": "FAST action tokenizer"
    },
    {
     "name": "libero",
     "url": "https://huggingface.co/datasets/physical-intelligence/libero",
     "kind": "dataset",
     "note": "LIBERO dataset"
    }
   ],
   "papers": [
    {
     "title": "π0: A Vision-Language-Action Flow Model for General Robot Control",
     "url": "https://arxiv.org/abs/2410.24164",
     "year": "2024"
    },
    {
     "title": "FAST: Efficient Action Tokenization for Vision-Language-Action Models",
     "url": "https://arxiv.org/abs/2501.09747",
     "year": "2025"
    },
    {
     "title": "π0.5: a Vision-Language-Action Model with Open-World Generalization",
     "url": "https://arxiv.org/abs/2504.16054",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "openpi",
     "url": "https://github.com/Physical-Intelligence/openpi",
     "note": "Open-source pi0 / pi0.5 models and code"
    },
    {
     "name": "real-time-chunking-kinetix",
     "url": "https://github.com/Physical-Intelligence/real-time-chunking-kinetix",
     "note": "Real-time action chunking experiments"
    },
    {
     "name": "aloha",
     "url": "https://github.com/Physical-Intelligence/aloha",
     "note": "ALOHA robot setup"
    }
   ],
   "openness": "openpi released under Apache-2.0 with open model checkpoints; some datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Karol Hausman",
     "role": "Co-founder and CEO",
     "url": "https://karolhausman.github.io/",
     "note": null
    },
    {
     "name": "Sergey Levine",
     "role": "Co-founder; professor at UC Berkeley",
     "url": "https://people.eecs.berkeley.edu/~svlevine/",
     "note": null
    },
    {
     "name": "Chelsea Finn",
     "role": "Co-founder; professor at Stanford",
     "url": "https://ai.stanford.edu/~cbfinn/",
     "note": null
    },
    {
     "name": "Brian Ichter",
     "role": "Co-founder",
     "url": "https://www.pi.website",
     "note": null
    },
    {
     "name": "Quan Vuong",
     "role": "Co-founder",
     "url": "https://www.pi.website",
     "note": null
    },
    {
     "name": "Lachy Groom",
     "role": "Co-founder",
     "url": "https://www.pi.website",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://github.com/Physical-Intelligence",
    "https://huggingface.co/physical-intelligence",
    "https://en.wikipedia.org/wiki/Physical_Intelligence_Inc."
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "physical-intelligence",
   "mlat": 37.79431,
   "mlon": -122.56922,
   "logo": true
  },
  {
   "orig_id": "pleias",
   "name": "Pleias",
   "alias": null,
   "full_name": "Pleias",
   "type": "startup",
   "founded": 2024,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://pleias.fr",
   "huggingface": "https://huggingface.co/PleIAs",
   "github": "https://github.com/Pleias",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/pleias",
   "contact_email": "contact@pleias.fr",
   "summary": "French research startup building language models trained on openly licensed and public-domain data. It maintains the Common Corpus dataset.",
   "tagline": "A Paris research company building language models and the Common Corpus dataset on openly licensed and rights-cleared data.",
   "about": [
    "Pleias is based at Station F in Paris and develops tools around training data: Common Corpus, an open collection of public-domain and permissively licensed text, Synth for synthetic training data, and Stratum for structuring documents.",
    "Common Corpus draws on government records, legal archives, scientific literature and multilingual sources, and is available on Hugging Face. The team has also trained small models, including a 600-million-parameter model for the Paris transport operator RATP.",
    "The company's authors include Pierre-Carl Langlais and Anastasia Stasenko, and it publishes research and models openly."
   ],
   "mission": "Pleias argues that capable models can be built on data with clear provenance and licensing. Its goal is to make that approach practical for industry and public institutions.",
   "notice": null,
   "known_for": [
    "Common Corpus open dataset",
    "Models trained on licensed data",
    "Synthetic data tooling",
    "European data provenance"
   ],
   "focus": [
    "datasets",
    "language models",
    "open data",
    "RAG"
   ],
   "contributions": [
    {
     "title": "Common Corpus",
     "note": "Released a large multilingual corpus of public-domain and openly licensed text.",
     "year": "2024"
    },
    {
     "title": "Small models in production",
     "note": "Delivered a 600M-parameter model for RATP's Sillon system.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "Quest-Best-Tokens",
     "url": "https://github.com/Pleias/Quest-Best-Tokens",
     "note": "An introduction to LLM Sampling (80 stars)"
    },
    {
     "name": "pleias_ScholasticAI",
     "url": "https://github.com/Pleias/pleias_ScholasticAI",
     "note": "Open-source repository (72 stars)"
    },
    {
     "name": "Pleias-RAG-Library",
     "url": "https://github.com/Pleias/Pleias-RAG-Library",
     "note": "Python library to use Pleias-RAG models (72 stars)"
    },
    {
     "name": "marginalia",
     "url": "https://github.com/Pleias/marginalia",
     "note": "Open-source repository (67 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "PleIAs/Monad",
     "url": "https://huggingface.co/PleIAs/Monad",
     "kind": "model",
     "note": "Model: 4,712 downloads, 72 likes."
    },
    {
     "name": "PleIAs/Baguettotron",
     "url": "https://huggingface.co/PleIAs/Baguettotron",
     "kind": "model",
     "note": "Model: 2,427 downloads, 270 likes."
    },
    {
     "name": "PleIAs/common_corpus",
     "url": "https://huggingface.co/datasets/PleIAs/common_corpus",
     "kind": "dataset",
     "note": "Dataset: 228,121 downloads, 452 likes."
    },
    {
     "name": "PleIAs/PleIAs-Editor",
     "url": "https://huggingface.co/spaces/PleIAs/PleIAs-Editor",
     "kind": "space",
     "note": "Space: 12 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Common Corpus: The Largest Collection of Ethical Data for LLM Pre-Training",
     "url": "https://arxiv.org/abs/2506.01732",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Common Corpus",
     "url": "https://huggingface.co/datasets/PleIAs/common_corpus",
     "note": "Large multilingual corpus of permissively licensed text"
    },
    {
     "name": "Pleias-RAG-1B",
     "url": "https://huggingface.co/PleIAs/Pleias-RAG-1B",
     "note": "Small model for retrieval-augmented generation"
    },
    {
     "name": "SYNTH",
     "url": "https://huggingface.co/datasets/PleIAs/SYNTH",
     "note": "Synthetic training dataset"
    },
    {
     "name": "Baguettotron",
     "url": "https://huggingface.co/PleIAs/Baguettotron",
     "note": "Small language model"
    }
   ],
   "openness": "Releases models and datasets on Hugging Face, emphasising training on openly licensed data; check each model card for its licence.",
   "key_people": [
    {
     "name": "Pierre-Carl Langlais",
     "role": "Co-founder and researcher",
     "url": "https://x.com/Dorialexander",
     "note": null
    },
    {
     "name": "Anastasia Stasenko",
     "role": "Researcher and author",
     "url": "https://pleias.ai/",
     "note": null
    },
    {
     "name": "Carlos Rosas",
     "role": "Author",
     "url": "https://pleias.ai/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact address printed on the website.",
   "sources": [
    "https://pleias.fr",
    "https://huggingface.co/PleIAs",
    "https://github.com/Pleias",
    "https://arxiv.org/abs/2506.01732"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "pleias",
   "mlat": 48.84709,
   "mlon": 2.1962,
   "logo": true
  },
  {
   "orig_id": "pluralis-research",
   "name": "Pluralis Research",
   "alias": null,
   "full_name": "Pluralis Research",
   "type": "startup",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://pluralis.ai",
   "huggingface": "https://huggingface.co/PluralisResearch",
   "github": "https://github.com/PluralisResearch",
   "x": "https://x.com/Pluralis",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Pluralis Research works on Protocol Learning, decentralized model-parallel training. Its GitHub has open training libraries; its Hugging Face org has no public models. Location not verified.",
   "tagline": "A research lab developing Protocol Learning, where large models are trained across many untrusted participants without any one holding the full weights.",
   "about": [
    "Pluralis Research was founded by Alexander Long, an ML PhD who previously worked in Amazon's international machine learning group. The lab is backed by investors including Union Square Ventures and CoinFund.",
    "Its foundational paper, Protocol Learning, Decentralized Frontier Risk and the No-Off Problem, appeared in December 2024. The lab then published work on model-parallel training over low-bandwidth links, including Subspace Networks and Unextractable Protocol Models at NeurIPS 2025.",
    "Code libraries are on GitHub, and its Agora technical report describes internet-scale pretraining across permissionless participants."
   ],
   "mission": "Pluralis argues that frontier models should be trainable by open networks rather than by a few firms, and works on the compression and parallelism needed.",
   "notice": null,
   "known_for": [
    "Protocol Learning concept",
    "Low-bandwidth model-parallel training",
    "Unextractable protocol models"
   ],
   "focus": [
    "distributed",
    "training"
   ],
   "contributions": [
    {
     "title": "Protocol Learning paper",
     "note": "Long posted the position paper on arXiv in December.",
     "year": "2024"
    },
    {
     "title": "Nesterov method for asynchronous pipeline parallel optimization",
     "note": "Paper accepted at ICML 2025.",
     "year": "2025"
    },
    {
     "title": "Subspace Networks",
     "note": "Communication-efficient model parallelism paper at NeurIPS 2025.",
     "year": "2025"
    },
    {
     "title": "Unextractable Protocol Models",
     "note": "Paper on training and inference without weight materialization, NeurIPS 2025.",
     "year": "2025"
    },
    {
     "title": "Agora technical report",
     "note": "Report on collective, permissionless pretraining of LLMs.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "node0",
     "url": "https://github.com/PluralisResearch/node0",
     "note": "Node0."
    },
    {
     "name": "agora",
     "url": "https://github.com/PluralisResearch/agora",
     "note": "Collaborative training library."
    },
    {
     "name": "AsyncPP",
     "url": "https://github.com/PluralisResearch/AsyncPP",
     "note": "Async pipeline parallel."
    },
    {
     "name": "stoa",
     "url": "https://github.com/PluralisResearch/stoa",
     "note": "RL post-training."
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism",
     "url": "https://arxiv.org/abs/2506.01260",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "node0",
     "url": "https://github.com/PluralisResearch/node0",
     "note": "Collaborative training"
    },
    {
     "name": "agora",
     "url": "https://github.com/PluralisResearch/agora",
     "note": "Collaborative training library"
    }
   ],
   "openness": "Open code on GitHub; no public weights seen.",
   "key_people": [
    {
     "name": "Alexander Long",
     "role": "Founder",
     "url": "https://pluralis.ai/",
     "note": null
    },
    {
     "name": "Thalaiyasingam Ajanthan",
     "role": "Researcher, paper author",
     "url": "https://pluralis.ai/",
     "note": null
    },
    {
     "name": "Sameera Ramasinghe",
     "role": "Researcher, paper author",
     "url": "https://pluralis.ai/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://pluralis.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://pluralis.ai",
    "https://github.com/PluralisResearch",
    "https://arxiv.org/abs/2505.01099"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "pluralis-research",
   "mlat": 37.7684,
   "mlon": -122.57099,
   "logo": true
  },
  {
   "orig_id": "polygl0t",
   "name": "Polygl0t",
   "alias": "Tucano",
   "full_name": "Polygl0t (Tucano)",
   "type": "academic",
   "founded": 2026,
   "country": "Germany",
   "city": "Bonn",
   "lat": 50.7374,
   "lon": 7.0982,
   "website": "https://github.com/Polygl0t",
   "huggingface": "https://huggingface.co/Polygl0t",
   "github": "https://github.com/Polygl0t",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Polygl0t develops open foundation models for low-resource languages, best known for the Tucano family of Portuguese models. It publishes weights, corpora such as GigaVerbo, and training code; the project is linked to the University of Bonn.",
   "tagline": "A University of Bonn project building open foundation models for low-resource languages, best known for the Tucano Portuguese models and the GigaVerbo corpus.",
   "about": [
    "Polygl0t, also referred to as Project Polyglot, develops efficient language models for low-resource languages. It is funded by the German BMBF and the North Rhine-Westphalia ministry as part of TRA Sustainable Futures at the University of Bonn, and trains on Bonn's Marvin cluster and Jülich's JUPITER Booster.",
    "Its Tucano models, first described in 2024, were trained on GigaVerbo, a corpus of about 200 billion deduplicated Portuguese tokens. The follow-up Tucano 2 paper of March 2026 introduces GigaVerbo-v2, with about 320 billion tokens, and models from roughly 0.5B to 3.7B parameters.",
    "The group publishes weights, datasets and training code, including an LLM foundry repository and a fork of the evaluation harness."
   ],
   "mission": "Polygl0t aims to close the linguistic divide in NLP by making capable open models for languages that large developers serve poorly.",
   "notice": null,
   "known_for": [
    "Tucano Portuguese models",
    "GigaVerbo corpus",
    "Open training code",
    "Low-resource language models"
   ],
   "focus": [
    "language models",
    "low-resource languages",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Tucano and GigaVerbo",
     "note": "Released Portuguese decoder models and a 200B-token corpus.",
     "year": "2024"
    },
    {
     "title": "Patterns publication",
     "note": "The Tucano work was published in the journal Patterns.",
     "year": "2025"
    },
    {
     "title": "Tucano 2",
     "note": "Released a second model family with GigaVerbo-v2 and post-training datasets.",
     "year": "2026"
    },
    {
     "title": "Workshop materials",
     "note": "Published workshop slides and exercises on GitHub.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "llm-foundry",
     "url": "https://github.com/Polygl0t/llm-foundry",
     "note": "Training codebase."
    },
    {
     "name": "workshop-slides",
     "url": "https://github.com/Polygl0t/workshop-slides",
     "note": "Workshop slides."
    },
    {
     "name": "workshop-exercises",
     "url": "https://github.com/Polygl0t/workshop-exercises",
     "note": "Exercises."
    }
   ],
   "hf_highlights": [
    {
     "name": "Tucano2-qwen-3.7B-Instruct",
     "url": "https://huggingface.co/Polygl0t/Tucano2-qwen-3.7B-Instruct",
     "kind": "model",
     "note": "Tucano2."
    },
    {
     "name": "Tucano2-qwen-1.5B-Instruct",
     "url": "https://huggingface.co/Polygl0t/Tucano2-qwen-1.5B-Instruct",
     "kind": "model",
     "note": "Tucano2 1.5B."
    },
    {
     "name": "gigaverbo-v2",
     "url": "https://huggingface.co/datasets/Polygl0t/gigaverbo-v2",
     "kind": "dataset",
     "note": "GigaVerbo v2."
    },
    {
     "name": "gigalekh-v1",
     "url": "https://huggingface.co/datasets/Polygl0t/gigalekh-v1",
     "kind": "dataset",
     "note": "Gigalekh."
    }
   ],
   "papers": [
    {
     "title": "Tucano: Advancing Neural Text Generation for Portuguese",
     "url": "https://arxiv.org/abs/2411.07854",
     "year": "2024"
    },
    {
     "title": "Tucano 2 Cool: Better Open Source LLMs for Portuguese",
     "url": "https://arxiv.org/abs/2603.03543",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "Tucano2-qwen-3.7B-Instruct",
     "url": "https://huggingface.co/Polygl0t/Tucano2-qwen-3.7B-Instruct",
     "note": "Portuguese instruct model"
    },
    {
     "name": "GigaVerbo v2",
     "url": "https://huggingface.co/datasets/Polygl0t/gigaverbo-v2",
     "note": "Portuguese corpus"
    },
    {
     "name": "llm-foundry",
     "url": "https://github.com/Polygl0t/llm-foundry",
     "note": "Training and eval codebase"
    }
   ],
   "openness": "Weights (Apache 2.0 on the instruct model checked), datasets and code are public on Hugging Face and GitHub. Brazilian focus but based in Germany.",
   "key_people": [
    {
     "name": "Nicholas Kluge Corrêa",
     "role": "Lead author of Tucano",
     "url": "https://github.com/Nkluge-correa",
     "note": null
    },
    {
     "name": "Aniket Sen",
     "role": "Co-author",
     "url": "https://github.com/aniketsen",
     "note": null
    },
    {
     "name": "Shiza Fatimah",
     "role": "Co-author",
     "url": "https://github.com/shiza-fatimah",
     "note": null
    },
    {
     "name": "Sophia Falk",
     "role": "Co-author",
     "url": "https://arxiv.org/abs/2411.07854",
     "note": null
    },
    {
     "name": "Lucie Flek",
     "role": "Co-author of Tucano 2",
     "url": "https://arxiv.org/abs/2603.03543",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://github.com/Polygl0t); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co/Polygl0t",
    "https://github.com/Polygl0t",
    "https://arxiv.org/abs/2411.07854"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "polygl0t",
   "logo": true
  },
  {
   "orig_id": "polymathic-ai",
   "name": "Polymathic AI",
   "alias": null,
   "full_name": "Polymathic AI",
   "type": "academic",
   "founded": 2023,
   "country": "United States",
   "city": "New York",
   "lat": 40.7128,
   "lon": -74.006,
   "website": "https://polymathic-ai.org",
   "huggingface": "https://huggingface.co/polymathic-ai",
   "github": "https://github.com/PolymathicAI",
   "x": "https://x.com/PolymathicAI",
   "discord": null,
   "linkedin": null,
   "contact_email": "info@polymathic-ai.org",
   "summary": "Simons Foundation and Flatiron Institute backed collaboration building open foundation models and datasets for scientific data, notably physics simulation and astronomy.",
   "tagline": "A Simons Foundation and Flatiron Institute backed collaboration building open foundation models and datasets for scientific data.",
   "about": [
    "Polymathic AI is led by principal investigator Shirley Ho and brings together researchers working across astrophysics, physics and machine learning. Its team includes Kyunghyun Cho, Miles Cranmer, Francois Lanusse and Michael McCabe.",
    "The group builds machine learning models that share concepts across scientific disciplines and releases them for researchers worldwide. A scientific advisory group includes Yann LeCun, Leslie Greengard and David Spergel.",
    "Recent news on its site covers agent systems for NMR structure elucidation and protein design."
   ],
   "mission": "Polymathic AI wants to give scientists pretrained models that carry across fields, so new problems start from shared knowledge rather than from scratch.",
   "notice": null,
   "known_for": [
    "Foundation models for science",
    "The Well physics dataset",
    "Astronomy foundation models",
    "Open scientific ML"
   ],
   "focus": [
    "scientific foundation models",
    "physics simulation",
    "astronomy",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Project site launched",
     "note": "Public project site went live.",
     "year": "2023"
    },
    {
     "title": "The Well released",
     "note": "A 15 TB collection of 16 physics simulation datasets, presented at NeurIPS 2024.",
     "year": "2024"
    },
    {
     "title": "Multimodal Universe",
     "note": "An open astrophysics dataset released alongside The Well in December 2024.",
     "year": "2024"
    },
    {
     "title": "Walrus",
     "note": "A 1.3B-parameter foundation model for continuum dynamics, trained on The Well and other data, described on arXiv in November 2025.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "the_well",
     "url": "https://github.com/PolymathicAI/the_well",
     "note": "15TB physics simulation datasets"
    },
    {
     "name": "walrus",
     "url": "https://github.com/PolymathicAI/walrus",
     "note": "Physical dynamics foundation model"
    },
    {
     "name": "multiple_physics_pretraining",
     "url": "https://github.com/PolymathicAI/multiple_physics_pretraining",
     "note": "Multiple physics pretraining code"
    },
    {
     "name": "AstroCLIP",
     "url": "https://github.com/PolymathicAI/AstroCLIP",
     "note": "Cross-modal galaxy model"
    }
   ],
   "hf_highlights": [
    {
     "name": "aion-base",
     "url": "https://huggingface.co/polymathic-ai/aion-base",
     "kind": "model",
     "note": "AION-1 base"
    },
    {
     "name": "turbulent_radiative_layer_2D",
     "url": "https://huggingface.co/datasets/polymathic-ai/turbulent_radiative_layer_2D",
     "kind": "dataset",
     "note": "The Well dataset"
    },
    {
     "name": "shear_flow",
     "url": "https://huggingface.co/datasets/polymathic-ai/shear_flow",
     "kind": "dataset",
     "note": "The Well dataset"
    }
   ],
   "papers": [
    {
     "title": "The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning",
     "url": "https://arxiv.org/abs/2412.00568",
     "year": "2024"
    },
    {
     "title": "Multiple Physics Pretraining for Physical Surrogate Models",
     "url": "https://arxiv.org/abs/2310.02994",
     "year": "2023"
    },
    {
     "title": "AstroCLIP: A Cross-Modal Foundation Model for Galaxies",
     "url": "https://arxiv.org/abs/2310.03024",
     "year": "2023"
    },
    {
     "title": "Walrus: A Cross-Domain Foundation Model for Continuum Dynamics",
     "url": "https://arxiv.org/abs/2511.15684",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "the_well",
     "url": "https://github.com/PolymathicAI/the_well",
     "note": "15TB physics simulation datasets"
    },
    {
     "name": "walrus",
     "url": "https://github.com/PolymathicAI/walrus",
     "note": "Physical dynamics foundation model"
    },
    {
     "name": "multiple_physics_pretraining",
     "url": "https://github.com/PolymathicAI/multiple_physics_pretraining",
     "note": "Multiple physics pretraining code"
    }
   ],
   "openness": "Open code (BSD-3 and MIT) and datasets/models on Hugging Face.",
   "key_people": [
    {
     "name": "Shirley Ho",
     "role": "Principal Investigator; Flatiron Institute",
     "url": "https://polymathic-ai.org/members/shirley-ho/",
     "note": "Leads the collaboration and is senior author of The Well."
    },
    {
     "name": "Kyunghyun Cho",
     "role": "Team member",
     "url": "https://kyunghyuncho.me/",
     "note": "Listed on the team page."
    },
    {
     "name": "Miles Cranmer",
     "role": "Team member",
     "url": "https://astroautomata.com/",
     "note": "Co-author of The Well."
    },
    {
     "name": "Francois Lanusse",
     "role": "Team member",
     "url": "https://flanusse.net/",
     "note": "Works on astronomy foundation models."
    },
    {
     "name": "Michael McCabe",
     "role": "Team member",
     "url": "https://polymathic-ai.org/members/michael-mccabe/",
     "note": "Second author of The Well."
    },
    {
     "name": "Rudy Morel",
     "role": "Team member",
     "url": "https://polymathic-ai.org/members/rudy-morel/",
     "note": "Co-author of The Well."
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on website.",
   "sources": [
    "https://polymathic-ai.org",
    "https://github.com/PolymathicAI",
    "https://huggingface.co/polymathic-ai",
    "https://www.simonsfoundation.org/2024/12/02/new-datasets-will-train-ai-models-to-think-like-scientists/"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "polymathic-ai",
   "mlat": 40.69553,
   "mlon": -74.07614,
   "logo": true
  },
  {
   "orig_id": "preferred-networks",
   "name": "Preferred Networks",
   "alias": null,
   "full_name": "Preferred Networks",
   "type": "startup",
   "founded": 2014,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://www.preferred.jp/en/",
   "huggingface": "https://huggingface.co/pfnet",
   "github": "https://github.com/pfnet",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Japanese deep learning company that publishes the PLaMo model family and related code.",
   "tagline": "A Tokyo deep learning company behind Chainer, the MN-Core accelerator and the PLaMo Japanese language models.",
   "about": [
    "Preferred Networks was established in March 2014 as a spin-off from Preferred Infrastructure, founded in 2006. Toru Nishikawa is CEO and Daisuke Okanohara is COO.",
    "In June 2015 the company released Chainer, an open-source deep learning framework that influenced dynamic-graph design; development ended in December 2019 as the team moved to PyTorch. It announced the MN-Core accelerator in December 2018, and the MN-3 system topped the Green500 list in June 2020.",
    "Preferred Networks continues to publish open code and Japanese language models, including the PLaMo family."
   ],
   "mission": "The company aims to apply deep learning to real-world problems, combining software, custom hardware and large-scale computing.",
   "notice": null,
   "known_for": [
    "Chainer deep learning framework",
    "MN-Core accelerator",
    "PLaMo Japanese models",
    "Green500 leading MN-3"
   ],
   "focus": [
    "language models",
    "Japanese",
    "finance"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Spun off from Preferred Infrastructure.",
     "year": "2014"
    },
    {
     "title": "Chainer released",
     "note": "Open-sourced the Chainer deep learning framework.",
     "year": "2015"
    },
    {
     "title": "MN-Core announced",
     "note": "Announced a dedicated deep learning accelerator.",
     "year": "2018"
    },
    {
     "title": "Chainer moves to PyTorch",
     "note": "Ended Chainer development and joined the PyTorch ecosystem.",
     "year": "2019"
    },
    {
     "title": "MN-3 tops Green500",
     "note": "MN-3 ranked first on the Green500 list.",
     "year": "2020"
    },
    {
     "title": "Optuna beta",
     "note": "Hyperparameter optimization framework released in December 2018.",
     "year": "2018"
    }
   ],
   "github_highlights": [
    {
     "name": "PaintsChainer",
     "url": "https://github.com/pfnet/PaintsChainer",
     "note": "Line drawing colorization."
    },
    {
     "name": "pfrl",
     "url": "https://github.com/pfnet/pfrl",
     "note": "PyTorch deep reinforcement learning library."
    },
    {
     "name": "pysen",
     "url": "https://github.com/pfnet/pysen",
     "note": "Python linting tool."
    },
    {
     "name": "plamo-translate-cli",
     "url": "https://github.com/pfnet/plamo-translate-cli",
     "note": "CLI for local translation with plamo-3-translate."
    }
   ],
   "hf_highlights": [
    {
     "name": "plamo-embedding-1b",
     "url": "https://huggingface.co/pfnet/plamo-embedding-1b",
     "kind": "model",
     "note": "Embedding model."
    },
    {
     "name": "plamo-3-translate-0.8b-GGUF",
     "url": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF",
     "kind": "model",
     "note": "Small translation model."
    },
    {
     "name": "plamo-2.1-2b-vl",
     "url": "https://huggingface.co/pfnet/plamo-2.1-2b-vl",
     "kind": "model",
     "note": "Vision-language model."
    },
    {
     "name": "bbh-ja",
     "url": "https://huggingface.co/datasets/pfnet/bbh-ja",
     "kind": "dataset",
     "note": "Japanese BIG-Bench Hard."
    }
   ],
   "papers": [
    {
     "title": "PLaMo-100B: A Ground-Up Language Model Designed for Japanese Proficiency",
     "url": "https://arxiv.org/abs/2410.07563",
     "year": "2024"
    },
    {
     "title": "Optuna: A Next-generation Hyperparameter Optimization Framework",
     "url": "https://arxiv.org/abs/1907.10902",
     "year": "2019"
    }
   ],
   "notable_releases": [
    {
     "name": "plamo-3-8b-fin-instruct",
     "url": "https://huggingface.co/pfnet/plamo-3-8b-fin-instruct",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "plamo-3-translate-8b-W4A16",
     "url": "https://huggingface.co/pfnet/plamo-3-translate-8b-W4A16",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face; licences vary per model.",
   "key_people": [
    {
     "name": "Toru Nishikawa",
     "role": "Co-founder and CEO",
     "url": "https://www.preferred.jp/en/",
     "note": null
    },
    {
     "name": "Daisuke Okanohara",
     "role": "Co-founder and COO",
     "url": "https://www.preferred.jp/en/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the website, GitHub or Hugging Face pages.",
   "sources": [
    "https://huggingface.co/pfnet",
    "https://www.preferred.jp/en/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "preferred-networks",
   "mlat": 35.71062,
   "mlon": 139.75529,
   "logo": true
  },
  {
   "orig_id": "prime-intellect",
   "name": "Prime Intellect",
   "alias": null,
   "full_name": "Prime Intellect",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://primeintellect.ai",
   "huggingface": "https://huggingface.co/PrimeIntellect",
   "github": "https://github.com/PrimeIntellect-ai",
   "x": "https://x.com/primeintellect",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Prime Intellect is a decentralized AI lab known for globally distributed training and RL tools, releasing models and code openly.",
   "tagline": "A lab and compute platform working on distributed training, asynchronous reinforcement learning and open environments for agentic models.",
   "about": [
    "Prime Intellect was co-founded by Vincent Weisser and Johannes Hagemann. Its 2024 launch post framed the company around open, decentralized AI development.",
    "In late 2024 it released INTELLECT-1, a 10B-class model trained on 1 trillion tokens across nodes on three continents with its PRIME distributed training framework, together with data and checkpoints. The company now offers inference, GPU compute, sandboxes and a model factory.",
    "Its open-source projects include prime-rl, a framework for asynchronous reinforcement learning at scale, and the Verifiers library with an Environments Hub that lists thousands of community environments."
   ],
   "mission": "Prime Intellect pursues open, globally distributed AI development and wants individuals and small teams to be able to train and own capable models.",
   "notice": null,
   "known_for": [
    "INTELLECT-1 distributed training",
    "prime-rl asynchronous RL",
    "Environments Hub",
    "Open decentralized AI"
   ],
   "focus": [
    "distributed",
    "language models",
    "reinforcement learning",
    "tooling"
   ],
   "contributions": [
    {
     "title": "Seed round",
     "note": "$5.5 million seed announced in April 2024.",
     "year": "2024"
    },
    {
     "title": "INTELLECT-1",
     "note": "Decentralized training of a 10B-parameter model; technical report on arXiv in December 2024.",
     "year": "2024"
    },
    {
     "title": "Series funding and prime-rl",
     "note": "A $15 million round led by Founders Fund in February 2025; asynchronous RL framework prime-rl open-sourced.",
     "year": "2025"
    },
    {
     "title": "INTELLECT-3",
     "note": "A 106B mixture-of-experts model trained with RL on 512 H200 GPUs; code and environments open-sourced in November 2025.",
     "year": "2025"
    },
    {
     "title": "$130 million Series A",
     "note": "Reported at a $1 billion valuation in July 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "prime-agent",
     "url": "https://github.com/PrimeIntellect-ai/prime-agent",
     "note": "Self-improving agent for coding workflows; most-starred repo."
    },
    {
     "name": "verifiers",
     "url": "https://github.com/PrimeIntellect-ai/verifiers",
     "note": "Library for RL environments and evals."
    },
    {
     "name": "prime-rl",
     "url": "https://github.com/PrimeIntellect-ai/prime-rl",
     "note": "Agentic RL training at scale."
    },
    {
     "name": "prime-diloco",
     "url": "https://github.com/PrimeIntellect-ai/prime-diloco",
     "note": "Framework for globally distributed training."
    }
   ],
   "hf_highlights": [
    {
     "name": "Qwen3-1.7B-Wordle-SFT",
     "url": "https://huggingface.co/PrimeIntellect/Qwen3-1.7B-Wordle-SFT",
     "kind": "model",
     "note": "SFT checkpoint for Wordle environment."
    },
    {
     "name": "Reverse-Text-RL",
     "url": "https://huggingface.co/datasets/PrimeIntellect/Reverse-Text-RL",
     "kind": "dataset",
     "note": "RL dataset for reverse-text environment."
    },
    {
     "name": "Multi-SWE-bench",
     "url": "https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-bench",
     "kind": "dataset",
     "note": "Multi-language SWE-bench data."
    }
   ],
   "papers": [
    {
     "title": "INTELLECT-1 Technical Report",
     "url": "https://arxiv.org/abs/2412.01152",
     "year": "2024"
    },
    {
     "title": "INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning",
     "url": "https://arxiv.org/abs/2505.07291",
     "year": "2025"
    },
    {
     "title": "OpenDiLoCo: An Open-Source Framework for Globally Distributed Low-Communication Training",
     "url": "https://arxiv.org/abs/2407.07852",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "INTELLECT-1",
     "url": "https://huggingface.co/PrimeIntellect/INTELLECT-1-fp32",
     "note": "Decentralised trained model"
    },
    {
     "name": "prime-rl",
     "url": "https://github.com/PrimeIntellect-ai/prime-rl",
     "note": "RL training"
    },
    {
     "name": "verifiers",
     "url": "https://github.com/PrimeIntellect-ai/verifiers",
     "note": "RL environments"
    }
   ],
   "openness": "Open weights, datasets and code.",
   "key_people": [
    {
     "name": "Vincent Weisser",
     "role": "Co-founder and CEO",
     "url": "https://www.ai.engineer/speakers/vincent-weisser",
     "note": null
    },
    {
     "name": "Johannes Hagemann",
     "role": "Co-founder",
     "url": "https://arxiv.org/abs/2412.01152",
     "note": null
    },
    {
     "name": "Sami Jaghouar",
     "role": "First author, INTELLECT-1",
     "url": "https://arxiv.org/abs/2412.01152",
     "note": "Led the INTELLECT-1 technical report."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://primeintellect.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://huggingface.co/PrimeIntellect",
    "https://github.com/PrimeIntellect-ai",
    "https://arxiv.org/abs/2412.01152"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "prime-intellect",
   "mlat": 37.7428,
   "mlon": -122.56568,
   "logo": true
  },
  {
   "orig_id": "princeton-nlp",
   "name": "Princeton NLP Group",
   "alias": null,
   "full_name": "Princeton NLP Group",
   "type": "academic",
   "founded": null,
   "country": "United States",
   "city": "Princeton",
   "lat": 40.3573,
   "lon": -74.6672,
   "website": "https://nlp.cs.princeton.edu",
   "huggingface": "https://huggingface.co/princeton-nlp",
   "github": "https://github.com/princeton-nlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Academic group releasing open benchmarks, methods and models for language models, including SWE-bench, SimPO and Tree of Thoughts.",
   "tagline": "The Princeton NLP group, source of SWE-bench, SimPO and Tree of Thoughts, with many open benchmarks, methods and models.",
   "about": [
    "Princeton NLP is the natural language processing research group at Princeton University. Its GitHub organization hosts widely used research code and its Hugging Face organization hosts models and datasets.",
    "The group's SWE-bench benchmark, published at ICLR 2024, tests whether language models can resolve real GitHub issues and became a standard evaluation for coding agents. SimPO, presented at NeurIPS 2024, is a reference-free preference optimization method, and Tree of Thoughts (NeurIPS 2023) is a deliberate problem-solving framework for language models.",
    "Karthik Narasimhan is among the authors of SWE-bench; Danqi Chen also leads work in the group."
   ],
   "mission": "The group studies how language models learn, reason and act, and releases the resulting benchmarks, methods and code so that other researchers can reproduce and extend them.",
   "notice": null,
   "known_for": [
    "SWE-bench",
    "SimPO preference optimization",
    "Tree of Thoughts",
    "Open academic NLP research"
   ],
   "focus": [
    "benchmarks",
    "alignment",
    "agents",
    "long context"
   ],
   "contributions": [
    {
     "title": "Tree of Thoughts",
     "note": "Introduced deliberate problem solving with language models.",
     "year": "2023"
    },
    {
     "title": "SWE-bench",
     "note": "Released the benchmark of 2,294 real GitHub issue tasks.",
     "year": "2023"
    },
    {
     "title": "SimPO",
     "note": "Released a simple reference-free preference optimization method with models.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "tree-of-thought-llm",
     "url": "https://github.com/princeton-nlp/tree-of-thought-llm",
     "note": "Tree of Thoughts"
    },
    {
     "name": "SimCSE",
     "url": "https://github.com/princeton-nlp/SimCSE",
     "note": "Contrastive sentence embeddings"
    },
    {
     "name": "MeZO",
     "url": "https://github.com/princeton-nlp/MeZO",
     "note": "Memory-efficient fine-tuning"
    },
    {
     "name": "SimPO",
     "url": "https://github.com/princeton-nlp/SimPO",
     "note": "Simple preference optimisation"
    }
   ],
   "hf_highlights": [
    {
     "name": "SWE-bench_Verified",
     "url": "https://huggingface.co/datasets/princeton-nlp/SWE-bench_Verified",
     "kind": "dataset",
     "note": "Verified SWE-bench subset"
    },
    {
     "name": "SWE-bench",
     "url": "https://huggingface.co/datasets/princeton-nlp/SWE-bench",
     "kind": "dataset",
     "note": "SWE-bench"
    },
    {
     "name": "Llama-3-8B-ProLong-512k-Base",
     "url": "https://huggingface.co/princeton-nlp/Llama-3-8B-ProLong-512k-Base",
     "kind": "model",
     "note": "Long-context model"
    }
   ],
   "papers": [
    {
     "title": "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?",
     "url": "https://arxiv.org/abs/2310.06770",
     "year": "2023"
    },
    {
     "title": "SimPO: Simple Preference Optimization with a Reference-Free Reward",
     "url": "https://arxiv.org/abs/2405.14734",
     "year": "2024"
    },
    {
     "title": "Tree of Thoughts: Deliberate Problem Solving with Large Language Models",
     "url": "https://arxiv.org/abs/2305.10601",
     "year": "2023"
    },
    {
     "title": "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering",
     "url": "https://arxiv.org/abs/2405.15793",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "tree-of-thought-llm",
     "url": "https://github.com/princeton-nlp/tree-of-thought-llm",
     "note": "Tree of Thoughts"
    },
    {
     "name": "SimCSE",
     "url": "https://github.com/princeton-nlp/SimCSE",
     "note": "Contrastive sentence embeddings"
    },
    {
     "name": "MeZO",
     "url": "https://github.com/princeton-nlp/MeZO",
     "note": "Memory-efficient fine-tuning"
    }
   ],
   "openness": "MIT-licensed code; benchmarks and models on Hugging Face.",
   "key_people": [
    {
     "name": "Karthik Narasimhan",
     "role": "Associate professor, co-author of SWE-bench and Tree of Thoughts",
     "url": "https://www.cs.princeton.edu/~karthikn/",
     "note": "Leads work on language agents."
    },
    {
     "name": "Danqi Chen",
     "role": "Associate professor, co-leader of the Princeton NLP Group",
     "url": "https://www.cs.princeton.edu/people/profile/danqic",
     "note": "Also associate director of Princeton Language and Intelligence."
    },
    {
     "name": "Shunyu Yao",
     "role": "First author of Tree of Thoughts, formerly a Princeton PhD student",
     "url": "https://ysymyth.github.io",
     "note": "Now a researcher at OpenAI."
    },
    {
     "name": "John Yang",
     "role": "Creator of SWE-bench and SWE-agent",
     "url": "https://john-b-yang.github.io",
     "note": "Earlier work done at Princeton; now a PhD student at Stanford."
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://nlp.cs.princeton.edu",
    "https://github.com/princeton-nlp",
    "https://huggingface.co/princeton-nlp"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "princeton-nlp-group",
   "logo": true
  },
  {
   "orig_id": "prior-labs",
   "name": "Prior Labs",
   "alias": null,
   "full_name": "Prior Labs",
   "type": "startup",
   "founded": 2023,
   "country": "Germany",
   "city": "Freiburg",
   "lat": 47.999,
   "lon": 7.8421,
   "website": "https://priorlabs.ai",
   "huggingface": "https://huggingface.co/Prior-Labs",
   "github": "https://github.com/PriorLabs",
   "x": "https://x.com/prior_labs",
   "discord": "https://discord.com/invite/VJRuU3bSxt",
   "linkedin": "https://www.linkedin.com/company/prior-labs",
   "contact_email": "sales@priorlabs.ai",
   "summary": "Startup building tabular foundation models. It releases the TabPFN family of models and open-source code.",
   "tagline": "A Freiburg company behind TabPFN, the tabular foundation model published in Nature that predicts on structured data without task-specific training.",
   "about": [
    "Prior Labs was founded by Frank Hutter, Noah Hollmann and Sauraj Gambhir to commercialize TabPFN, a transformer pretrained on synthetic data that handles tabular prediction in a single forward pass.",
    "TabPFN was published in Nature in early 2025. The company raised a 9 million euro pre-seed round led by Balderton in February 2025, and its scientific advisory board includes Yann LeCun and Bernhard Schölkopf.",
    "In May 2026 SAP announced plans to acquire Prior Labs, with commitments that it remains a separate legal entity and keeps its open-source work."
   ],
   "mission": "Prior Labs wants foundation models to become the default tool for tabular data, which dominates business and science data.",
   "notice": "SAP announced in May 2026 that it intends to acquire Prior Labs. Prior Labs has said it will remain a separate entity and keep its open-source commitments.",
   "known_for": [
    "TabPFN tabular foundation model",
    "Nature publication",
    "Open-source tabular AI",
    "Single-pass prediction"
   ],
   "focus": [
    "tabular data",
    "foundation models",
    "time series"
   ],
   "contributions": [
    {
     "title": "TabPFN in Nature",
     "note": "The TabPFN paper appeared in Nature.",
     "year": "2025"
    },
    {
     "title": "Pre-seed round",
     "note": "Raised 9 million euros led by Balderton.",
     "year": "2025"
    },
    {
     "title": "TabPFN-2.5",
     "note": "Extended the model to larger tables of up to 50,000 samples and 2,000 features.",
     "year": "2025"
    },
    {
     "title": "SAP acquisition announced",
     "note": "SAP announced it would acquire the company.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "TabPFN",
     "url": "https://github.com/PriorLabs/TabPFN",
     "note": "Foundation model for tabular data (8,477 stars)"
    },
    {
     "name": "tabpfn-time-series",
     "url": "https://github.com/PriorLabs/tabpfn-time-series",
     "note": "Zero-shot time series forecasting with TabPFN (479 stars)"
    },
    {
     "name": "tabpfn-extensions",
     "url": "https://github.com/PriorLabs/tabpfn-extensions",
     "note": "Community extensions for TabPFN (344 stars)"
    },
    {
     "name": "tabpfn-client",
     "url": "https://github.com/PriorLabs/tabpfn-client",
     "note": "Easy API access to TabPFN (256 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "Prior-Labs/TabPFN-v2-reg",
     "url": "https://huggingface.co/Prior-Labs/TabPFN-v2-reg",
     "kind": "model",
     "note": "TabPFN-v2 regression weights."
    },
    {
     "name": "Prior-Labs/TabPFN-v2-clf",
     "url": "https://huggingface.co/Prior-Labs/TabPFN-v2-clf",
     "kind": "model",
     "note": "TabPFN-v2 classification weights."
    },
    {
     "name": "Prior-Labs/tabpfn_3_5",
     "url": "https://huggingface.co/Prior-Labs/tabpfn_3_5",
     "kind": "model",
     "note": "TabPFN 3.5 weights."
    },
    {
     "name": "Prior-Labs/tabpfn_2_5",
     "url": "https://huggingface.co/Prior-Labs/tabpfn_2_5",
     "kind": "model",
     "note": "TabPFN 2.5 weights."
    }
   ],
   "papers": [
    {
     "title": "TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second",
     "url": "https://arxiv.org/abs/2207.01848",
     "year": "2022"
    },
    {
     "title": "Accurate predictions on small data with a tabular foundation model",
     "url": "https://www.nature.com/articles/s41586-024-08328-6",
     "year": "2025"
    },
    {
     "title": "TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models",
     "url": "https://arxiv.org/abs/2511.08667",
     "year": "2025"
    },
    {
     "title": "From Tables to Time: Extending TabPFN-v2 to Time Series Forecasting",
     "url": "https://arxiv.org/abs/2501.02945",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "TabPFN",
     "url": "https://github.com/PriorLabs/TabPFN",
     "note": "Tabular foundation model code"
    },
    {
     "name": "tabpfn_2_5",
     "url": "https://huggingface.co/Prior-Labs/tabpfn_2_5",
     "note": "TabPFN model weights"
    },
    {
     "name": "tabpfn-time-series",
     "url": "https://github.com/PriorLabs/tabpfn-time-series",
     "note": "Zero-shot time series forecasting"
    }
   ],
   "openness": "Code is Apache-2.0; weights are public on Hugging Face (check each card for licence).",
   "key_people": [
    {
     "name": "Frank Hutter",
     "role": "Co-founder; professor at University of Freiburg",
     "url": "https://ml.informatik.uni-freiburg.de/profile/hutter/",
     "note": null
    },
    {
     "name": "Noah Hollmann",
     "role": "Co-founder",
     "url": "https://priorlabs.ai",
     "note": null
    },
    {
     "name": "Sauraj Gambhir",
     "role": "Co-founder",
     "url": "https://priorlabs.ai",
     "note": null
    }
   ],
   "discord_note": "Community server linked from Prior Labs' contact page for discussion of TabPFN.",
   "contact_note": "Its contact page has a form and a sales address, plus the community Discord.",
   "sources": [
    "https://priorlabs.ai",
    "https://huggingface.co/Prior-Labs",
    "https://github.com/PriorLabs",
    "https://arxiv.org/abs/2207.01848"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "prior-labs",
   "mlat": 48.02065,
   "mlon": 7.78606,
   "logo": true
  },
  {
   "orig_id": "probabl",
   "name": "probabl",
   "alias": null,
   "full_name": "probabl",
   "type": "startup",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://probabl.ai",
   "huggingface": "https://huggingface.co/probabl-ai",
   "github": "https://github.com/probabl-ai",
   "x": "https://x.com/probabl_ai",
   "discord": "https://discord.com/invite/f3j8hBSWM7",
   "linkedin": "https://www.linkedin.com/company/probabl",
   "contact_email": null,
   "summary": "Paris company building open-source tools for tabular data science, including the skore library.",
   "tagline": "A Paris mission-driven company spun out of Inria to sustain scikit-learn and the open-source data science ecosystem around it.",
   "about": [
    "Probabl was formed in 2023 and launched publicly in early 2024 to give the scikit-learn ecosystem a sustainable institutional home. Scikit-learn core maintainers who had been employed by Inria joined as co-founders, and Yann Lechelle, formerly of Scaleway, became chief executive.",
    "Inria and the French state are shareholders, and the company has a mission-driven legal structure. Its stated aim is the sustainable development of a data science commons, with professional services and products built around open-source libraries such as scikit-learn and skrub.",
    "In October 2025 Sifted reported a 13 million euro seed round. In January 2026 Gaël Varoquaux, a scikit-learn co-founder and Inria researcher, announced he was taking the role of chief science officer."
   ],
   "mission": "Probabl intends to make open-source machine learning tools sustainable and to offer professional services and products around them.",
   "notice": null,
   "known_for": [
    "scikit-learn stewardship",
    "Inria spin-off",
    "Mission-driven open-source company",
    "Tabular data science tools"
   ],
   "focus": [
    "tabular data",
    "tooling",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Company launch",
     "note": "Launched as a spin-off around scikit-learn and Inria.",
     "year": "2024"
    },
    {
     "title": "Seed round",
     "note": "Reported 13 million euro seed round.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "skore",
     "url": "https://github.com/probabl-ai/skore",
     "note": "Open-source library for tracking and evaluating data science work (684 stars)"
    },
    {
     "name": "skore-skills",
     "url": "https://github.com/probabl-ai/skore-skills",
     "note": "Tabular data science skills for guardrailing AI agents (146 stars)"
    },
    {
     "name": "youtube-appendix",
     "url": "https://github.com/probabl-ai/youtube-appendix",
     "note": "Resources for its educational content (50 stars)"
    },
    {
     "name": "sklearn-ragger-duck",
     "url": "https://github.com/probabl-ai/sklearn-ragger-duck",
     "note": "Proof-of-concept RAG over the scikit-learn docs (29 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "probabl-ai/ScikitLLM-Model",
     "url": "https://huggingface.co/probabl-ai/ScikitLLM-Model",
     "kind": "model",
     "note": "Model for scikit-learn code assistance."
    },
    {
     "name": "probabl-ai/ScikitLLM-Model-exl2",
     "url": "https://huggingface.co/probabl-ai/ScikitLLM-Model-exl2",
     "kind": "model",
     "note": "Quantised variant of ScikitLLM."
    },
    {
     "name": "probabl-ai/LLM_article",
     "url": "https://huggingface.co/datasets/probabl-ai/LLM_article",
     "kind": "dataset",
     "note": "Dataset from an LLM article."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "skore",
     "url": "https://github.com/probabl-ai/skore",
     "note": "Library for tracking data science work"
    },
    {
     "name": "skore-skills",
     "url": "https://github.com/probabl-ai/skore-skills",
     "note": "Skills for guardrailing AI agents"
    },
    {
     "name": "youtube-appendix",
     "url": "https://github.com/probabl-ai/youtube-appendix",
     "note": "Educational resources"
    }
   ],
   "openness": "Libraries are open-source under MIT or BSD licences.",
   "key_people": [
    {
     "name": "Yann Lechelle",
     "role": "Co-founder and CEO",
     "url": "https://probabl.ai",
     "note": "Previously ran Scaleway and was an entrepreneur in residence at Inria."
    },
    {
     "name": "Gaël Varoquaux",
     "role": "Co-founder and chief science officer",
     "url": "https://gael-varoquaux.info/",
     "note": "Inria researcher and scikit-learn co-founder; also works on skrub."
    }
   ],
   "discord_note": "Community server linked from probabl's contact page.",
   "contact_note": "Its contact page has a form to talk to an expert and links to the community Discord.",
   "sources": [
    "https://probabl.ai",
    "https://huggingface.co/probabl-ai",
    "https://github.com/probabl-ai"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "probabl",
   "mlat": 48.81065,
   "mlon": 2.21196,
   "logo": true
  },
  {
   "orig_id": "projecte-aina",
   "name": "Projecte Aina",
   "alias": null,
   "full_name": "Projecte Aina",
   "type": "institute",
   "founded": 2021,
   "country": "Spain",
   "city": "Barcelona",
   "lat": 41.3874,
   "lon": 2.1686,
   "website": "https://projecteaina.cat",
   "huggingface": "https://huggingface.co/projecte-aina",
   "github": "https://github.com/projecte-aina",
   "x": "https://x.com/projecte_aina",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/projecte-aina",
   "contact_email": null,
   "summary": "Catalan government-backed initiative, run with Barcelona Supercomputing Center, that publishes open Catalan language models, speech models and datasets.",
   "tagline": "A Catalan government initiative led by the Barcelona Supercomputing Center that builds open Catalan language models, speech resources and corpora.",
   "about": [
    "Projecte Aina is a Catalan government initiative created by the Generalitat de Catalunya's digital policies department to ensure that Catalan is well supported in language technology. The Barcelona Supercomputing Center (BSC) carries out the technical work, with its Language Technologies Unit credited as author of the models.",
    "The project has built a large Catalan text corpus, started a Catalan speech corpus and published models and datasets on Hugging Face under permissive licenses. Its Salamandra family of 2B, 7B and 40B models, released under Apache 2.0, was trained on the MareNostrum 5 supercomputer with Aina funding.",
    "Resources are intended for reuse by companies, public bodies and researchers, and the project also runs community programs such as the Aina Challenge."
   ],
   "mission": "Aina aims to ensure that technology understands and speaks Catalan, with open resources that others can reuse.",
   "notice": null,
   "known_for": [
    "Catalan language corpora",
    "Open Catalan models",
    "Salamandra models on MareNostrum 5",
    "Government-backed language technology"
   ],
   "focus": [
    "Catalan",
    "language models",
    "speech",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Project begins",
     "note": "BSC starts building Catalan corpora and models for the Generalitat de Catalunya.",
     "year": "2022"
    },
    {
     "title": "Catalan models on Hugging Face",
     "note": "Open Catalan language, speech and translation models and datasets are published under the projecte-aina organization.",
     "year": "2023"
    },
    {
     "title": "Salamandra",
     "note": "Released 2B, 7B and 40B multilingual models under Apache 2.0, with Catalan upsampled in the training data.",
     "year": "2024"
    },
    {
     "title": "Aina Challenge and territorial intelligence programs",
     "note": "Programs run in parallel with the model work to encourage reuse of the resources.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "lm-catalan",
     "url": "https://github.com/projecte-aina/lm-catalan",
     "note": "Catalan language models and resources from the Aina project (27 stars)"
    },
    {
     "name": "oTranscribe-plus",
     "url": "https://github.com/projecte-aina/oTranscribe-plus",
     "note": "Audio transcription tool with offline ASR support, a fork (26 stars)"
    },
    {
     "name": "tts-api",
     "url": "https://github.com/projecte-aina/tts-api",
     "note": "REST API for synthesising Catalan speech (17 stars)"
    },
    {
     "name": "spacy",
     "url": "https://github.com/projecte-aina/spacy",
     "note": "Pre-production spaCy releases for Catalan (13 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "projecte-aina/roberta-base-ca-v2-cased-ner",
     "url": "https://huggingface.co/projecte-aina/roberta-base-ca-v2-cased-ner",
     "kind": "model",
     "note": "Catalan named entity recognition model."
    },
    {
     "name": "projecte-aina/FLOR-760M",
     "url": "https://huggingface.co/projecte-aina/FLOR-760M",
     "kind": "model",
     "note": "Catalan-focused language model."
    },
    {
     "name": "projecte-aina/CATalog",
     "url": "https://huggingface.co/datasets/projecte-aina/CATalog",
     "kind": "dataset",
     "note": "Catalan text corpus."
    },
    {
     "name": "projecte-aina/matxa-alvocat-tts-ca",
     "url": "https://huggingface.co/spaces/projecte-aina/matxa-alvocat-tts-ca",
     "kind": "space",
     "note": "Catalan text-to-speech demo."
    }
   ],
   "papers": [
    {
     "title": "Are Multilingual Models the Best Choice for Moderately Under-resourced Languages? A Comprehensive Assessment for Catalan",
     "url": "https://arxiv.org/abs/2107.07903",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Aguila-7B",
     "url": "https://huggingface.co/projecte-aina/aguila-7b",
     "note": "Catalan-Spanish-English model"
    },
    {
     "name": "FLOR-6.3B",
     "url": "https://huggingface.co/projecte-aina/FLOR-6.3B",
     "note": "Catalan-adapted language model"
    },
    {
     "name": "CATalog",
     "url": "https://huggingface.co/datasets/projecte-aina/CATalog",
     "note": "Catalan corpus"
    },
    {
     "name": "lm-catalan",
     "url": "https://github.com/projecte-aina/lm-catalan",
     "note": "Catalan language model resources"
    }
   ],
   "openness": "Models, datasets and code are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Barcelona Supercomputing Center",
     "role": "Lead institution",
     "url": "https://www.bsc.es/",
     "note": null
    },
    {
     "name": "Marta Villegas",
     "role": "Author of an article on AINA",
     "url": "https://traces.uab.cat/record/133338?ln=en",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The Aina site has a contact page for the initiative.",
   "sources": [
    "https://projecteaina.cat",
    "https://huggingface.co/projecte-aina",
    "https://github.com/projecte-aina"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "projecte-aina",
   "mlat": 41.4374,
   "mlon": 2.1686,
   "logo": true
  },
  {
   "orig_id": "proxecto-nos",
   "name": "Proxecto Nos",
   "alias": "University of Santiago de Compostela",
   "full_name": "Proxecto Nos (University of Santiago de Compostela)",
   "type": "academic",
   "founded": 2020,
   "country": "Spain",
   "city": "Santiago de Compostela",
   "lat": 42.8782,
   "lon": -8.5448,
   "website": "https://nos.gal",
   "huggingface": "https://huggingface.co/proxectonos",
   "github": "https://github.com/proxectonos",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Galician language technology project at Universidade de Santiago de Compostela releasing open Galician language models, corpora and speech tools.",
   "tagline": "The Xunta de Galicia project, run by USC, that creates open Galician language models, corpora and speech tools.",
   "about": [
    "Proxecto Nós was launched in 2020 by the Xunta de Galicia and entrusted to the University of Santiago de Compostela through CiTIUS and the Instituto da Lingua Galega (ILG). It aims to give Galician openly licensed resources and tools in intelligent technologies.",
    "It produced CorpusNós, described as the largest openly available collection of Galician text, and the first generative language models focused on Galician, including the 1.3 billion parameter Carballo, adapted from Catalan-Spanish-English models of the Aina project. A 2024 arXiv paper describes the models.",
    "The project also operates public demonstrators for machine translation, speech synthesis, speech recognition and retrieval-augmented chat. CiTIUS was later selected to lead the Galician model in a Microsoft collaboration with BSC and HiTZ on models for Spain's official languages."
   ],
   "mission": "The project seeks to ensure Galician has the digital resources and AI technology needed for everyday use.",
   "notice": null,
   "known_for": [
    "CorpusNOS Galician text corpus",
    "First Galician generative LLMs",
    "Open Galician speech tools"
   ],
   "focus": [
    "Galician",
    "language models",
    "speech"
   ],
   "contributions": [
    {
     "title": "Proxecto Nós starts",
     "note": "Launched by the Xunta de Galicia and hosted by CiTIUS and the ILG at USC.",
     "year": "2020"
    },
    {
     "title": "Project paper at TDLE",
     "note": "Presented the project's aims and resources.",
     "year": "2022"
    },
    {
     "title": "First Galician LLMs",
     "note": "Released Carballo models trained on CorpusNós, as described in the arXiv paper.",
     "year": "2024"
    },
    {
     "title": "Microsoft collaboration",
     "note": "CiTIUS selected to lead the Galician model in a multilingual initiative with BSC and HiTZ.",
     "year": "2024"
    },
    {
     "title": "Galician evaluation tool presented",
     "note": "A tool for evaluating Galician in large AI models was presented at the European Parliament.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "language-models",
     "url": "https://github.com/proxectonos/language-models",
     "note": "Modelos de lingua para o galego / Language models for Galician (7 stars)"
    },
    {
     "name": "corpora",
     "url": "https://github.com/proxectonos/corpora",
     "note": "Corpus de texto ou voz en galego (monolingües e multilingües) para diferentes tarefas PLN/ Monolingual and mul (6 stars)"
    },
    {
     "name": "pipeline",
     "url": "https://github.com/proxectonos/pipeline",
     "note": "Open-source repository. (3 stars)"
    },
    {
     "name": "demos",
     "url": "https://github.com/proxectonos/demos",
     "note": "Nesta sección pode acceder ás diferentes demonstracións do proxecto Nós.  (2 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "proxectonos/Carballo-bloom-1.3B",
     "url": "https://huggingface.co/proxectonos/Carballo-bloom-1.3B",
     "kind": "model",
     "note": "Model with 7 likes on the Hub."
    },
    {
     "name": "proxectonos/Llama-3.1-Carballo",
     "url": "https://huggingface.co/proxectonos/Llama-3.1-Carballo",
     "kind": "model",
     "note": "Model with 6 likes on the Hub."
    },
    {
     "name": "proxectonos/Nos_Parlaspeech-GL",
     "url": "https://huggingface.co/datasets/proxectonos/Nos_Parlaspeech-GL",
     "kind": "dataset",
     "note": "Dataset with 2 likes on the Hub."
    },
    {
     "name": "proxectonos/Nos_LLM_galician",
     "url": "https://huggingface.co/spaces/proxectonos/Nos_LLM_galician",
     "kind": "space",
     "note": "Space with 1 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Open Generative Large Language Models for Galician",
     "url": "https://arxiv.org/abs/2406.13893",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Carballo-bloom-1.3B",
     "url": "https://huggingface.co/proxectonos/Carballo-bloom-1.3B",
     "note": "Galician language model"
    },
    {
     "name": "Llama-3.1-Carballo",
     "url": "https://huggingface.co/proxectonos/Llama-3.1-Carballo",
     "note": "Galician-adapted Llama"
    },
    {
     "name": "Nos TTS",
     "url": "https://huggingface.co/proxectonos/Nos_TTS-celtia-vits-graphemes",
     "note": "Galician text to speech"
    },
    {
     "name": "language-models repository",
     "url": "https://github.com/proxectonos/language-models",
     "note": "Galician language models"
    }
   ],
   "openness": "Models, speech tools and corpora are public on Hugging Face and GitHub.",
   "key_people": [
    {
     "name": "Pablo Gamallo",
     "role": "CiTIUS researcher, lead author of the Galician LLM paper",
     "url": "https://arxiv.org/abs/2406.13893",
     "note": "Linked page is the paper listing the project team."
    },
    {
     "name": "José Ramom Pichel",
     "role": "Researcher, co-author of the Galician LLM paper",
     "url": "https://arxiv.org/abs/2406.13893",
     "note": "Linked page is the paper listing the project team."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the project website.",
   "sources": [
    "https://nos.gal",
    "https://huggingface.co/proxectonos",
    "https://github.com/proxectonos"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "proxecto-nos",
   "logo": true
  },
  {
   "orig_id": "pruna-ai",
   "name": "Pruna AI",
   "alias": null,
   "full_name": "Pruna AI",
   "type": "startup",
   "founded": 2023,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.pruna.ai",
   "huggingface": "https://huggingface.co/PrunaAI",
   "github": "https://github.com/PrunaAI",
   "x": "https://x.com/PrunaAI",
   "discord": "https://discord.gg/JFQmtFKCjd",
   "linkedin": "https://www.linkedin.com/company/pruna-ai",
   "contact_email": "hello@pruna.ai",
   "summary": "Startup working on model efficiency: compression, quantisation and compilation. It open-sources the pruna library and publishes optimised models.",
   "tagline": "A model efficiency company that open-sources the pruna framework for compressing and accelerating AI models.",
   "about": [
    "Pruna AI was founded in 2023 by Bertrand Charpentier, Rayan Nait Mazi, John Rachwan and Stephan Günnemann, researchers in AI efficiency with ties to the Technical University of Munich. It operates research hubs in Munich and Paris.",
    "In November 2024 it announced a 6.5 million dollar seed round led by EQT Ventures, with Daphni, Motier and Kima Ventures participating. In March 2025 it open-sourced its optimization framework, which combines caching, pruning, quantization and distillation and checks quality loss.",
    "The pruna package is MIT-licensed, a paid tier adds further methods, and the company publishes many compressed models on Hugging Face."
   ],
   "mission": "Pruna aims to make AI models smaller, faster and cheaper to run.",
   "notice": null,
   "known_for": [
    "pruna optimization framework",
    "Model compression",
    "Compressed models on Hugging Face",
    "Inference speedups"
   ],
   "focus": [
    "efficiency",
    "quantisation",
    "image generation",
    "tooling"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Started by TUM-linked efficiency researchers in Munich and Paris.",
     "year": "2023"
    },
    {
     "title": "Seed round",
     "note": "6.5 million dollar round led by EQT Ventures.",
     "year": "2024"
    },
    {
     "title": "Framework open-sourced",
     "note": "Released the pruna optimization framework under the MIT license.",
     "year": "2025"
    },
    {
     "title": "Compressed models on Hugging Face",
     "note": "Publishes large numbers of optimized open models under the PrunaAI organization.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "pruna",
     "url": "https://github.com/PrunaAI/pruna",
     "note": "Pruna is a model optimization framework built for developers, enabling you to deliver faster, more efficient m (1,391 stars)"
    },
    {
     "name": "awesome-ai-efficiency",
     "url": "https://github.com/PrunaAI/awesome-ai-efficiency",
     "note": "A curated list of materials on AI efficiency (240 stars)"
    },
    {
     "name": "ai-efficiency-courses",
     "url": "https://github.com/PrunaAI/ai-efficiency-courses",
     "note": "Courses on building, compressing, evaluating, and deploying efficient AI models. (92 stars)"
    },
    {
     "name": "ComfyUI_pruna",
     "url": "https://github.com/PrunaAI/ComfyUI_pruna",
     "note": "This is a ComfyUI node that integrates pruna (67 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "PrunaAI/P-H3-Image-Comfy",
     "url": "https://huggingface.co/PrunaAI/P-H3-Image-Comfy",
     "kind": "model",
     "note": "Model: 1,200,813 downloads, 467 likes."
    },
    {
     "name": "PrunaAI/P-Q-Image-2.1",
     "url": "https://huggingface.co/PrunaAI/P-Q-Image-2.1",
     "kind": "model",
     "note": "Model: 24,238 downloads, 194 likes."
    },
    {
     "name": "PrunaAI/media-assets",
     "url": "https://huggingface.co/datasets/PrunaAI/media-assets",
     "kind": "dataset",
     "note": "Dataset: 39,092 downloads, 1 likes."
    },
    {
     "name": "PrunaAI/P-H3-Image",
     "url": "https://huggingface.co/spaces/PrunaAI/P-H3-Image",
     "kind": "space",
     "note": "Space: 175 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "pruna",
     "url": "https://github.com/PrunaAI/pruna",
     "note": "Model optimisation framework"
    },
    {
     "name": "P-H3-Image",
     "url": "https://huggingface.co/PrunaAI/P-H3-Image",
     "note": "Image generation model"
    },
    {
     "name": "awesome-ai-efficiency",
     "url": "https://github.com/PrunaAI/awesome-ai-efficiency",
     "note": "Curated list of efficiency resources"
    }
   ],
   "openness": "The pruna framework is Apache-2.0; many compressed models are published on Hugging Face.",
   "key_people": [
    {
     "name": "Bertrand Charpentier",
     "role": "Co-founder, president and chief scientist",
     "url": "https://www.pruna.ai/about",
     "note": null
    },
    {
     "name": "Rayan Nait Mazi",
     "role": "Co-founder and CEO",
     "url": "https://www.pruna.ai/about",
     "note": null
    },
    {
     "name": "John Rachwan",
     "role": "Co-founder and CTO",
     "url": "https://www.pruna.ai/about",
     "note": null
    },
    {
     "name": "Stephan Günnemann",
     "role": "Co-founder and chief strategy officer",
     "url": "https://www.pruna.ai/about",
     "note": "Professor at TUM."
    }
   ],
   "discord_note": "Community Discord for users of the pruna library.",
   "contact_note": "Email printed on the website; Discord.",
   "sources": [
    "https://www.pruna.ai",
    "https://huggingface.co/PrunaAI",
    "https://github.com/PrunaAI"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "pruna-ai",
   "mlat": 48.78042,
   "mlon": 2.24665,
   "logo": true
  },
  {
   "orig_id": "pyannote",
   "name": "pyannoteAI",
   "alias": null,
   "full_name": "pyannoteAI",
   "type": "startup",
   "founded": 2024,
   "country": "France",
   "city": "Paris",
   "lat": 48.8566,
   "lon": 2.3522,
   "website": "https://www.pyannote.ai",
   "huggingface": "https://huggingface.co/pyannote",
   "github": "https://github.com/pyannote",
   "x": "https://x.com/pyannoteAI",
   "discord": "https://discord.gg/vux8UH9QmV",
   "linkedin": "https://www.linkedin.com/company/pyannoteai",
   "contact_email": "support@pyannote.ai",
   "summary": "Company behind pyannote.audio, the open-source toolkit for speaker diarization. It also offers commercial models.",
   "tagline": "The creators of pyannote.audio, the open-source speaker diarization toolkit, now also offering hosted models through pyannoteAI.",
   "about": [
    "pyannote.audio was created by Hervé Bredin and collaborators, described in an ICASSP 2020 paper, and is released under the MIT licence. Its pretrained models are widely downloaded on Hugging Face.",
    "pyannoteAI was founded in 2024 with Vincent Molina as CEO and Bredin as chief science officer. It sells faster hosted models such as precision-2.",
    "In late 2025 the project released pyannote.audio 4.0 with the open community-1 model."
   ],
   "mission": "The team aims to give developers accurate speaker intelligence, keeping the core toolkit open while funding development through hosted models.",
   "notice": null,
   "known_for": [
    "pyannote.audio",
    "Speaker diarization",
    "community-1 open model",
    "Hugging Face pipelines"
   ],
   "focus": [
    "speech",
    "speaker diarization",
    "tooling"
   ],
   "contributions": [
    {
     "title": "pyannote.audio paper",
     "note": "Published the toolkit at ICASSP.",
     "year": "2020"
    },
    {
     "title": "pyannoteAI founded",
     "note": "Launched the company around the toolkit.",
     "year": "2024"
    },
    {
     "title": "pyannote.audio 4.0",
     "note": "Released version 4.0 with community-1.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "pyannote-audio",
     "url": "https://github.com/pyannote/pyannote-audio",
     "note": "Neural building blocks for speaker diarization (10,806 stars)"
    },
    {
     "name": "pyannote-video",
     "url": "https://github.com/pyannote/pyannote-video",
     "note": "Face detection, tracking and clustering in videos (477 stars)"
    },
    {
     "name": "pyannote-metrics",
     "url": "https://github.com/pyannote/pyannote-metrics",
     "note": "Toolkit for reproducible diarization evaluation (258 stars)"
    },
    {
     "name": "pyannote-core",
     "url": "https://github.com/pyannote/pyannote-core",
     "note": "Data structures for temporal segments with labels (126 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "pyannote/speaker-diarization-3.1",
     "url": "https://huggingface.co/pyannote/speaker-diarization-3.1",
     "kind": "model",
     "note": "Speaker diarization pipeline, the group's most downloaded model."
    },
    {
     "name": "pyannote/speaker-diarization-community-1",
     "url": "https://huggingface.co/pyannote/speaker-diarization-community-1",
     "kind": "model",
     "note": "Community diarization pipeline."
    },
    {
     "name": "pyannote/segmentation-3.0",
     "url": "https://huggingface.co/pyannote/segmentation-3.0",
     "kind": "model",
     "note": "Speaker segmentation model."
    },
    {
     "name": "pyannote/wespeaker-voxceleb-resnet34-LM",
     "url": "https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM",
     "kind": "model",
     "note": "Speaker embedding model."
    }
   ],
   "papers": [
    {
     "title": "pyannote.audio: neural building blocks for speaker diarization",
     "url": "https://arxiv.org/abs/1911.01255",
     "year": "2019"
    },
    {
     "title": "Powerset multi-class cross entropy loss for neural speaker diarization",
     "url": "https://arxiv.org/abs/2310.13025",
     "year": "2023"
    },
    {
     "title": "End-to-end speaker segmentation for overlap-aware resegmentation",
     "url": "https://arxiv.org/abs/2104.04045",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "pyannote-audio",
     "url": "https://github.com/pyannote/pyannote-audio",
     "note": "Neural building blocks for speaker diarization"
    },
    {
     "name": "speaker-diarization-3.1",
     "url": "https://huggingface.co/pyannote/speaker-diarization-3.1",
     "note": "Diarization pipeline"
    },
    {
     "name": "segmentation-3.0",
     "url": "https://huggingface.co/pyannote/segmentation-3.0",
     "note": "Segmentation model"
    }
   ],
   "openness": "Core code is MIT-licensed and several models are public on Hugging Face.",
   "key_people": [
    {
     "name": "Hervé Bredin",
     "role": "Creator of pyannote.audio and chief science officer",
     "url": "https://herve.niderb.fr/",
     "note": null
    },
    {
     "name": "Vincent Molina",
     "role": "CEO",
     "url": "https://www.pyannote.ai/about",
     "note": null
    },
    {
     "name": "Juan Coria",
     "role": "CTO (named by some sources)",
     "url": "https://www.pyannote.ai/about",
     "note": null
    }
   ],
   "discord_note": "Community server linked from the pyannoteAI site.",
   "contact_note": "The pyannoteAI contact page lists a support address and the community Discord.",
   "sources": [
    "https://www.pyannote.ai",
    "https://huggingface.co/pyannote",
    "https://github.com/pyannote",
    "https://arxiv.org/abs/1911.01255"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "pyannoteai",
   "mlat": 48.76048,
   "mlon": 2.29561,
   "logo": true
  },
  {
   "orig_id": "qcri",
   "name": "Qatar Computing Research Institute",
   "alias": null,
   "full_name": "Qatar Computing Research Institute",
   "type": "institute",
   "founded": null,
   "country": "Qatar",
   "city": "Doha",
   "lat": 25.29,
   "lon": 51.53,
   "website": "https://www.hbku.edu.qa/en/qcri",
   "huggingface": "https://huggingface.co/QCRI",
   "github": "https://github.com/qcri",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Research institute of HBKU focused on Arabic technology, multilingual NLP and responsible AI. Government-linked.",
   "tagline": "Qatar's national computing research institute at HBKU, developer of Arabic language technology including the Fanar generative AI platform.",
   "about": [
    "The Qatar Computing Research Institute is part of Hamad Bin Khalifa University and focuses on Arabic language technology, multilingual NLP and responsible AI. It is government-linked and not an independent lab.",
    "It developed Fanar, an Arabic-centric platform covering language, speech and image generation, sponsored by Qatar's Ministry of Communications and Information Technology. The core models, Fanar Star (7B) and Fanar Prime (9B, continued from Gemma-2), are described in a January 2025 technical report, and Fanar 2.0 followed.",
    "QCRI also released tools such as Farasa for Arabic processing and publishes research code and models on GitHub and Hugging Face."
   ],
   "mission": "QCRI works to advance Arabic AI and make language technology accessible for Arabic speakers.",
   "notice": null,
   "known_for": [
    "Fanar Arabic AI platform",
    "Farasa Arabic tools",
    "Arabic NLP research",
    "Open models on Hugging Face"
   ],
   "focus": [
    "Arabic",
    "NLP",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Farasa",
     "note": "Released an Arabic text processing toolkit.",
     "year": "2016"
    },
    {
     "title": "Fanar launched",
     "note": "Presented Qatar's sovereign Arabic LLM at World Summit AI in Doha.",
     "year": "2024"
    },
    {
     "title": "Fanar technical report",
     "note": "Published the Fanar report describing the multimodal platform.",
     "year": "2025"
    },
    {
     "title": "Fanar 2.0",
     "note": "HBKU released the enhanced platform.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "Arabesque",
     "url": "https://github.com/qcri/Arabesque",
     "note": "Scalable graph mining."
    },
    {
     "name": "NADEEF",
     "url": "https://github.com/qcri/NADEEF",
     "note": "Data cleaning system."
    },
    {
     "name": "dialectID",
     "url": "https://github.com/qcri/dialectID",
     "note": "Arabic dialect detection."
    },
    {
     "name": "DeepER",
     "url": "https://github.com/qcri/DeepER",
     "note": "Entity resolution."
    }
   ],
   "hf_highlights": [
    {
     "name": "QCRI/Fanar-1-9B-Instruct",
     "url": "https://huggingface.co/QCRI/Fanar-1-9B-Instruct",
     "kind": "model",
     "note": "Fanar 9B."
    },
    {
     "name": "QCRI/Fanar-2-27B-Instruct",
     "url": "https://huggingface.co/QCRI/Fanar-2-27B-Instruct",
     "kind": "model",
     "note": "Fanar 2 27B."
    },
    {
     "name": "QCRI/CrisisMMD",
     "url": "https://huggingface.co/datasets/QCRI/CrisisMMD",
     "kind": "dataset",
     "note": "Crisis dataset."
    },
    {
     "name": "QCRI/MemeLens",
     "url": "https://huggingface.co/datasets/QCRI/MemeLens",
     "kind": "dataset",
     "note": "Meme dataset."
    }
   ],
   "papers": [
    {
     "title": "Fanar: An Arabic-Centric Multimodal Generative AI Platform",
     "url": "https://arxiv.org/abs/2501.13944",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "LlamaLens",
     "url": "https://huggingface.co/QCRI/LlamaLens",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "MemeLens-VLM",
     "url": "https://huggingface.co/QCRI/MemeLens-VLM",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open models and datasets.",
   "key_people": [
    {
     "name": "Fahim Dalvi",
     "role": "Researcher working on large language models at QCRI",
     "url": "https://fdalvi.github.io",
     "note": "Writes on the Fanar releases."
    },
    {
     "name": "Firoj Alam",
     "role": "Senior scientist at QCRI",
     "url": "https://firojalam.one",
     "note": "Co-author of the Fanar report."
    },
    {
     "name": "Sanjay Chawla",
     "role": "Co-author of the Fanar report",
     "url": "https://arxiv.org/abs/2501.13944",
     "note": "Linked page is the Fanar paper."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact is via the HBKU QCRI page.",
   "sources": [
    "https://huggingface.co/QCRI",
    "https://www.hbku.edu.qa/en/qcri"
   ],
   "confidence": "medium",
   "region": "Middle East",
   "id": "qatar-computing-research-institute",
   "logo": true
  },
  {
   "orig_id": "reka-ai",
   "name": "Reka",
   "alias": null,
   "full_name": "Reka",
   "type": "startup",
   "founded": 2022,
   "country": "United States",
   "city": "Sunnyvale",
   "lat": 37.3688,
   "lon": -122.0363,
   "website": "https://reka.ai",
   "huggingface": "https://huggingface.co/RekaAI",
   "github": "https://github.com/reka-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "contact@reka.ai",
   "summary": "Multimodal AI startup that releases some open-weight models (Reka Flash 3) and benchmarks such as Vibe-Eval.",
   "tagline": "A multimodal AI company that releases selected open-weight models such as Reka Flash 3 and Reka Flash 3.1.",
   "about": [
    "Reka was founded in 2022 by researchers formerly at DeepMind and elsewhere in industry, and builds multimodal models. Dani Yogatama, also an associate professor at USC, is cofounder and CEO. Its flagship models are served through its own platform, while selected models are released openly.",
    "In March 2025 it released Reka Flash 3, a 21B reasoning model trained with reinforcement learning, under Apache 2.0. In July 2025 it open-sourced Reka Flash 3.1 together with the Reka Quant quantization method.",
    "It also publishes the Vibe-Eval benchmark for multimodal models, and a 2024 technical report describes its Core, Flash and Edge models."
   ],
   "mission": "Reka aims to build capable multimodal models and to share part of its work with the community.",
   "notice": null,
   "known_for": [
    "Reka Flash 3 open weights",
    "Reka Quant",
    "Vibe-Eval benchmark",
    "Multimodal models"
   ],
   "focus": [
    "multimodal",
    "language models",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Reka Flash 3",
     "note": "Released a 21B Apache-licensed reasoning model.",
     "year": "2025"
    },
    {
     "title": "Reka Flash 3.1 and Reka Quant",
     "note": "Open-sourced an updated model and a quantization method.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "reka-vibe-eval",
     "url": "https://github.com/reka-ai/reka-vibe-eval",
     "note": "Vibe-Eval benchmark."
    },
    {
     "name": "rekaquant",
     "url": "https://github.com/reka-ai/rekaquant",
     "note": "Quantisation."
    },
    {
     "name": "research-eval",
     "url": "https://github.com/reka-ai/research-eval",
     "note": "Search-augmented LLM benchmark."
    }
   ],
   "hf_highlights": [
    {
     "name": "reka-flash-3",
     "url": "https://huggingface.co/RekaAI/reka-flash-3",
     "kind": "model",
     "note": "Flash 3."
    },
    {
     "name": "RekaDaily-10k-raw",
     "url": "https://huggingface.co/datasets/RekaAI/RekaDaily-10k-raw",
     "kind": "dataset",
     "note": "Raw daily data."
    },
    {
     "name": "VibeEval",
     "url": "https://huggingface.co/datasets/RekaAI/VibeEval",
     "kind": "dataset",
     "note": "Vibe-Eval."
    }
   ],
   "papers": [
    {
     "title": "Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models",
     "url": "https://arxiv.org/abs/2404.12387",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "reka-flash-3",
     "url": "https://huggingface.co/RekaAI/reka-flash-3",
     "note": "Open reasoning model"
    },
    {
     "name": "reka-vibe-eval",
     "url": "https://github.com/reka-ai/reka-vibe-eval",
     "note": "Benchmark"
    }
   ],
   "openness": "Selected open weights and eval code; main models are closed.",
   "key_people": [
    {
     "name": "Dani Yogatama",
     "role": "Cofounder and CEO",
     "url": "https://dyogatama.github.io",
     "note": "Associate professor of computer science at USC."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://reka.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://reka.ai",
    "https://huggingface.co/RekaAI",
    "https://github.com/reka-ai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "reka",
   "logo": true
  },
  {
   "orig_id": "rinna",
   "name": "rinna",
   "alias": null,
   "full_name": "rinna",
   "type": "startup",
   "founded": 2020,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.66,
   "lon": 139.73,
   "website": "https://rinna.co.jp/",
   "huggingface": "https://huggingface.co/rinna",
   "github": "https://github.com/rinnakk",
   "x": "https://twitter.com/rinna_research",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Japanese generative AI company with a long record of releasing pre-trained Japanese models.",
   "tagline": "A Japanese generative AI company known for releasing pretrained Japanese GPT, BERT and speech models, including the Youri and Nekomata LLMs.",
   "about": [
    "rinna was established in 2020 when the team behind Microsoft's Japanese Rinna chatbot, launched in 2015, separated from Microsoft to operate independently.",
    "Its research group releases base models for Japanese text, audio and images, including GPT, BERT, HuBERT, CLIP and Stable Diffusion variants, on Hugging Face and GitHub. Its early Japanese GPT-2 models became common baselines for Japanese NLP.",
    "The Youri series continues pretraining of Llama 2 on about 40 billion tokens of mixed Japanese and English data, and the Nekomata series builds on Alibaba's Qwen models with a tokenizer that is more efficient for Japanese."
   ],
   "mission": "rinna aims to give Japanese developers open pretrained models for their language.",
   "notice": null,
   "known_for": [
    "Japanese GPT-2 models",
    "Youri models",
    "Nekomata models",
    "Japanese pretrained models"
   ],
   "focus": [
    "language models",
    "Japanese",
    "speech"
   ],
   "contributions": [
    {
     "title": "Company established",
     "note": "Rinna team spun out of Microsoft Japan.",
     "year": "2020"
    },
    {
     "title": "Japanese GPT-2",
     "note": "Released small, medium and xsmall Japanese GPT-2 models.",
     "year": "2021"
    },
    {
     "title": "Youri",
     "note": "Released Llama 2 based Japanese-English models.",
     "year": "2023"
    },
    {
     "title": "Nekomata",
     "note": "Released Qwen-based models with a Japanese-efficient tokenizer.",
     "year": "2023"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "japanese-gpt-neox-small",
     "url": "https://huggingface.co/rinna/japanese-gpt-neox-small",
     "kind": "model",
     "note": "Most downloaded model."
    },
    {
     "name": "japanese-clip-vit-b-16",
     "url": "https://huggingface.co/rinna/japanese-clip-vit-b-16",
     "kind": "model",
     "note": "Japanese CLIP."
    },
    {
     "name": "japanese-roberta-base",
     "url": "https://huggingface.co/rinna/japanese-roberta-base",
     "kind": "model",
     "note": "Japanese RoBERTa."
    },
    {
     "name": "japanese-gpt2-medium",
     "url": "https://huggingface.co/rinna/japanese-gpt2-medium",
     "kind": "model",
     "note": "Japanese GPT-2."
    }
   ],
   "papers": [
    {
     "title": "Release of Pre-Trained Models for the Japanese Language",
     "url": "https://arxiv.org/abs/2404.01657",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "qwen2.5-bakeneko-32b",
     "url": "https://huggingface.co/rinna/qwen2.5-bakeneko-32b",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "llama-3-youko-70b",
     "url": "https://huggingface.co/rinna/llama-3-youko-70b",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face and code on GitHub.",
   "key_people": [],
   "discord_note": null,
   "contact_note": "Contact via its website or Hugging Face page.",
   "sources": [
    "https://huggingface.co/rinna",
    "https://rinna.co.jp/"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "rinna",
   "logo": true
  },
  {
   "orig_id": "roboflow",
   "name": "Roboflow",
   "alias": null,
   "full_name": "Roboflow",
   "type": "startup",
   "founded": 2019,
   "country": "United States",
   "city": "Des Moines",
   "lat": 41.5868,
   "lon": -93.625,
   "website": "https://roboflow.com",
   "huggingface": "https://huggingface.co/Roboflow",
   "github": "https://github.com/roboflow",
   "x": "https://x.com/roboflow",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computer-vision platform company that open-sources supervision, RF-DETR and curated datasets.",
   "tagline": "A computer vision platform company that open-sources Supervision, Inference, RF-DETR and the Universe dataset hub.",
   "about": [
    "Roboflow was started by Joseph Nelson (CEO) and Brad Dwyer (CTO) after their 2017 Magic Sudoku augmented reality app. The company's own site says they launched it in 2020, while other sources date incorporation to 2019.",
    "Open-source projects include Supervision for vision utilities, Inference for running models, and RF-DETR, a real-time detector accepted at ICLR 2026. Roboflow Universe hosts shared datasets and models.",
    "RF-DETR authors include Peter Robicheaux, James Gallagher and Isaac Robinson alongside Nelson."
   ],
   "mission": "Roboflow wants to make computer vision accessible to every developer.",
   "notice": null,
   "known_for": [
    "Supervision library",
    "RF-DETR detector",
    "Roboflow Universe datasets",
    "Developer-friendly vision tools"
   ],
   "focus": [
    "computer vision",
    "object detection",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Magic Sudoku",
     "note": "Dwyer and Nelson built the AR app that preceded the company.",
     "year": "2017"
    },
    {
     "title": "Company founded",
     "note": "Incorporated in 2019 and launched through Y Combinator in 2020.",
     "year": "2019"
    },
    {
     "title": "Series B",
     "note": "40 million dollar round led by GV, announced by the company.",
     "year": "2024"
    },
    {
     "title": "RF-DETR",
     "note": "Released a real-time detection transformer; the paper lists Isaac Robinson, Peter Robicheaux, Matvei Popov, Deva Ramanan and Neehar Peri.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "supervision",
     "url": "https://github.com/roboflow/supervision",
     "note": "Reusable computer vision tools"
    },
    {
     "name": "notebooks",
     "url": "https://github.com/roboflow/notebooks",
     "note": "Computer vision tutorials"
    },
    {
     "name": "inference",
     "url": "https://github.com/roboflow/inference",
     "note": "Edge inference server"
    }
   ],
   "hf_highlights": [
    {
     "name": "rf-detr-base",
     "url": "https://huggingface.co/Roboflow/rf-detr-base",
     "kind": "model",
     "note": "RF-DETR base"
    },
    {
     "name": "rf-detr-large",
     "url": "https://huggingface.co/Roboflow/rf-detr-large",
     "kind": "model",
     "note": "RF-DETR large"
    }
   ],
   "papers": [
    {
     "title": "RF-DETR: Neural Architecture Search for Real-Time Detection Transformers",
     "url": "https://arxiv.org/abs/2511.09554",
     "year": "2025"
    },
    {
     "title": "Roboflow 100: A Rich, Multi-Domain Object Detection Benchmark",
     "url": "https://arxiv.org/abs/2211.13523",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "supervision",
     "url": "https://github.com/roboflow/supervision",
     "note": "Reusable computer vision tools"
    },
    {
     "name": "notebooks",
     "url": "https://github.com/roboflow/notebooks",
     "note": "Computer vision tutorials"
    },
    {
     "name": "inference",
     "url": "https://github.com/roboflow/inference",
     "note": "Edge inference server"
    }
   ],
   "openness": "MIT and Apache-2.0 libraries and models.",
   "key_people": [
    {
     "name": "Joseph Nelson",
     "role": "Co-founder and CEO",
     "url": "https://yespress.io/joseph-nelson",
     "note": null
    },
    {
     "name": "Brad Dwyer",
     "role": "Co-founder and CTO",
     "url": "https://roboflow.com/about",
     "note": null
    },
    {
     "name": "Peter Robicheaux",
     "role": "RF-DETR author",
     "url": "https://www.catalyzex.com/author/Peter%20Robicheaux",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://roboflow.com",
    "https://github.com/roboflow",
    "https://huggingface.co/Roboflow"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "roboflow",
   "logo": true
  },
  {
   "orig_id": "rwkv",
   "name": "RWKV",
   "alias": null,
   "full_name": "RWKV",
   "type": "collective",
   "founded": null,
   "country": "Hong Kong",
   "city": "Hong Kong",
   "lat": 22.32,
   "lon": 114.17,
   "website": "https://www.rwkv.com/",
   "huggingface": "https://huggingface.co/RWKV",
   "github": "https://github.com/RWKV",
   "x": "https://twitter.com/RWKV_AI",
   "discord": "https://discord.gg/bDSBUMeFpc",
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source community behind the RWKV RNN-style language model architecture, now a Linux Foundation AI project. Location is approximate.",
   "tagline": "An open-source community behind the RWKV recurrent language model architecture, hosted as a Linux Foundation AI & Data project.",
   "about": [
    "RWKV started as a project by independent researcher Bo Peng, who leads it, and grew into a community project under the Linux Foundation AI & Data.",
    "The architecture has gone through versions named Dove, Eagle, Finch and Goose. RWKV-7 Goose, described in a March 2025 arXiv paper, generalizes the delta rule with vector-valued gating and supports state tracking. Models of 0.4B, 1.5B and 2.9B were released under Apache 2.0.",
    "Contributors come from EleutherAI, Recursal AI and several universities."
   ],
   "mission": "RWKV aims to offer an efficient alternative to transformers with linear-time inference.",
   "notice": null,
   "known_for": [
    "RWKV architecture",
    "RWKV-7 Goose",
    "Linux Foundation project",
    "Efficient RNN-style LLMs"
   ],
   "focus": [
    "language models",
    "architecture",
    "training infrastructure"
   ],
   "contributions": [
    {
     "title": "RWKV-4",
     "note": "First publicly scaled RWKV models.",
     "year": "2023"
    },
    {
     "title": "Joins Linux Foundation",
     "note": "Project joined the Linux Foundation in September 2023.",
     "year": "2023"
    },
    {
     "title": "RWKV-7 Goose",
     "note": "Released March 2025 under Apache 2.0; the paper has 18 authors including Ruichong Zhang, Daniel Goldstein and Eric Alcaide.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "rwkv.cpp",
     "url": "https://github.com/RWKV/rwkv.cpp",
     "note": "CPU inference for RWKV."
    },
    {
     "name": "RWKV-infctx-trainer",
     "url": "https://github.com/RWKV/RWKV-infctx-trainer",
     "note": "Infinite context trainer."
    },
    {
     "name": "RWKV-LM",
     "url": "https://github.com/RWKV/RWKV-LM",
     "note": "Core RWKV repo."
    },
    {
     "name": "rwkv-onnx",
     "url": "https://github.com/RWKV/rwkv-onnx",
     "note": "ONNX converter."
    }
   ],
   "hf_highlights": [
    {
     "name": "RWKV/v5-Eagle-7B-HF",
     "url": "https://huggingface.co/RWKV/v5-Eagle-7B-HF",
     "kind": "model",
     "note": "Eagle 7B."
    },
    {
     "name": "RWKV/rwkv-4-169m-pile",
     "url": "https://huggingface.co/RWKV/rwkv-4-169m-pile",
     "kind": "model",
     "note": "Small Pile model."
    },
    {
     "name": "RWKV/RWKV7-G1j-2.9B-20260831",
     "url": "https://huggingface.co/RWKV/RWKV7-G1j-2.9B-20260831",
     "kind": "model",
     "note": "RWKV-7 2.9B."
    },
    {
     "name": "RWKV/EagleX-WorldContinued",
     "url": "https://huggingface.co/datasets/RWKV/EagleX-WorldContinued",
     "kind": "dataset",
     "note": "Training data."
    }
   ],
   "papers": [
    {
     "title": "RWKV-7 \"Goose\" with Expressive Dynamic State Evolution",
     "url": "https://arxiv.org/abs/2503.14456",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "RWKV7-G1l-13.3B-20261031",
     "url": "https://huggingface.co/RWKV/RWKV7-G1l-13.3B-20261031",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "RWKV7-G1l-7.2B-20261031",
     "url": "https://huggingface.co/RWKV/RWKV7-G1l-7.2B-20261031",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open weights and code (RWKV-LM, RWKV-Runner).",
   "key_people": [
    {
     "name": "Bo Peng",
     "role": "Creator and project lead",
     "url": "https://github.com/BlinkDL",
     "note": "First author of the RWKV-7 paper."
    }
   ],
   "discord_note": "Main community Discord linked from rwkv.com.",
   "contact_note": "rwkv.com links the community Discord and GitHub.",
   "sources": [
    "https://huggingface.co/RWKV",
    "https://www.rwkv.com/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "rwkv",
   "logo": true
  },
  {
   "orig_id": "sakana-ai",
   "name": "Sakana AI",
   "alias": null,
   "full_name": "Sakana AI",
   "type": "startup",
   "founded": 2023,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://sakana.ai",
   "huggingface": "https://huggingface.co/SakanaAI",
   "github": "https://github.com/SakanaAI",
   "x": "https://twitter.com/SakanaAILabs",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Tokyo R&D company pursuing nature-inspired foundation models. Releases research code, models and Japanese datasets.",
   "tagline": "A Tokyo research company founded by former Google researchers, known for evolutionary model merging and The AI Scientist.",
   "about": [
    "Sakana AI was founded in 2023 by David Ha (CEO), Llion Jones (CTO) and Ren Ito (COO). Ha and Jones had worked at Google, where they met in Tokyo.",
    "The lab pursues nature-inspired methods. Evolutionary Model Merge combines layers and weights of existing models without retraining, and The AI Scientist is a framework for automated scientific discovery.",
    "It releases research code, models and Japanese datasets."
   ],
   "mission": "Sakana AI aims to develop foundation models inspired by natural processes such as evolution and collective intelligence.",
   "notice": null,
   "known_for": [
    "Evolutionary Model Merge",
    "The AI Scientist",
    "Japanese open models",
    "Nature-inspired AI"
   ],
   "focus": [
    "language models",
    "evolutionary methods",
    "datasets",
    "Japanese"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Incorporated in Tokyo.",
     "year": "2023"
    },
    {
     "title": "Evolutionary Model Merge",
     "note": "Released open models from automated merging.",
     "year": "2024"
    },
    {
     "title": "The AI Scientist",
     "note": "Released the framework for automated research.",
     "year": "2024"
    },
    {
     "title": "Series B",
     "note": "Announced a Series B round.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "AI-Scientist",
     "url": "https://github.com/SakanaAI/AI-Scientist",
     "note": "Automated open-ended scientific discovery."
    },
    {
     "name": "AI-Scientist-v2",
     "url": "https://github.com/SakanaAI/AI-Scientist-v2",
     "note": "Agentic tree search version."
    },
    {
     "name": "continuous-thought-machines",
     "url": "https://github.com/SakanaAI/continuous-thought-machines",
     "note": "Continuous Thought Machines."
    },
    {
     "name": "ShinkaEvolve",
     "url": "https://github.com/SakanaAI/ShinkaEvolve",
     "note": "Sample-efficient program evolution."
    }
   ],
   "hf_highlights": [
    {
     "name": "TinySwallow-1.5B-Instruct",
     "url": "https://huggingface.co/SakanaAI/TinySwallow-1.5B-Instruct",
     "kind": "model",
     "note": "Small Japanese instruct model."
    },
    {
     "name": "TinySwallow-1.5B",
     "url": "https://huggingface.co/SakanaAI/TinySwallow-1.5B",
     "kind": "model",
     "note": "Base model."
    },
    {
     "name": "ALE-Bench",
     "url": "https://huggingface.co/datasets/SakanaAI/ALE-Bench",
     "kind": "dataset",
     "note": "Algorithm engineering benchmark."
    },
    {
     "name": "AI-CUDA-Engineer-Archive",
     "url": "https://huggingface.co/datasets/SakanaAI/AI-CUDA-Engineer-Archive",
     "kind": "dataset",
     "note": "Archive of generated CUDA kernels."
    }
   ],
   "papers": [
    {
     "title": "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery",
     "url": "https://arxiv.org/abs/2408.06292",
     "year": "2024"
    },
    {
     "title": "Evolutionary Optimization of Model Merging Recipes",
     "url": "https://arxiv.org/abs/2403.13187",
     "year": "2024"
    },
    {
     "title": "The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search",
     "url": "https://arxiv.org/abs/2504.08066",
     "year": "2025"
    },
    {
     "title": "Continuous Thought Machines",
     "url": "https://arxiv.org/abs/2505.05522",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "JFinBench",
     "url": "https://huggingface.co/datasets/SakanaAI/JFinBench",
     "note": "Dataset"
    },
    {
     "name": "EDINET-Bench",
     "url": "https://huggingface.co/datasets/SakanaAI/EDINET-Bench",
     "note": "Dataset"
    },
    {
     "name": "FishMath-SFT-Data",
     "url": "https://huggingface.co/datasets/SakanaAI/FishMath-SFT-Data",
     "note": "Dataset"
    }
   ],
   "openness": "Open weights, datasets and research code on Hugging Face and GitHub; some products closed.",
   "key_people": [
    {
     "name": "David Ha",
     "role": "Co-founder and CEO",
     "url": "https://x.com/hardmaru",
     "note": null
    },
    {
     "name": "Llion Jones",
     "role": "Co-founder and CTO",
     "url": "https://sakana.ai",
     "note": null
    },
    {
     "name": "Ren Ito",
     "role": "Co-founder and COO",
     "url": "https://sakana.ai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website, GitHub or Hugging Face pages.",
   "sources": [
    "https://huggingface.co/SakanaAI",
    "https://sakana.ai"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "sakana-ai",
   "mlat": 35.74124,
   "mlon": 139.69,
   "logo": true
  },
  {
   "orig_id": "sapienza-nlp",
   "name": "Sapienza NLP",
   "alias": "Minerva",
   "full_name": "Sapienza NLP (Minerva)",
   "type": "academic",
   "founded": null,
   "country": "Italy",
   "city": "Rome",
   "lat": 41.9028,
   "lon": 12.4964,
   "website": "https://nlp.uniroma1.it",
   "huggingface": "https://huggingface.co/sapienzanlp",
   "github": "https://github.com/SapienzaNLP",
   "x": "https://x.com/SapienzaNLP",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "NLP group at Sapienza University of Rome that leads the Minerva family of Italian-English open language models and publishes entity linking, coreference and semantics tools.",
   "tagline": "The Sapienza University of Rome NLP group led by Roberto Navigli, creator of BabelNet and the Minerva Italian language models.",
   "about": [
    "Sapienza NLP is headed by Roberto Navigli, a professor at Sapienza and Fellow of AAAI, ACL, EurAI and ELLIS. His ERC grants (MultiJEDI and MOUSSE) funded work on BabelNet and word sense disambiguation.",
    "In April 2024 the group released Minerva, a family of language models trained from scratch on Italian and English, each on more than 500 billion words. The work was led by Navigli with Edoardo Barba and Simone Conia within the FAIR project, with CINECA compute.",
    "The group publishes tools for entity linking, coreference and semantics."
   ],
   "mission": "The group studies multilingual meaning and builds open resources for Italian and other languages.",
   "notice": null,
   "known_for": [
    "Minerva Italian LLMs",
    "BabelNet",
    "Word sense disambiguation",
    "Entity linking tools"
   ],
   "focus": [
    "language models",
    "Italian",
    "semantics",
    "entity linking"
   ],
   "contributions": [
    {
     "title": "MultiJEDI ERC grant",
     "note": "Funded BabelNet development.",
     "year": "2011"
    },
    {
     "title": "Minerva",
     "note": "Released Italian-English models trained from scratch.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "relik",
     "url": "https://github.com/SapienzaNLP/relik",
     "note": "Retrieve, Read and LinK: Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget (ACL 2 (519 stars)"
    },
    {
     "name": "extend",
     "url": "https://github.com/SapienzaNLP/extend",
     "note": "Entity Disambiguation as text extraction (ACL 2022) (182 stars)"
    },
    {
     "name": "spring",
     "url": "https://github.com/SapienzaNLP/spring",
     "note": "SPRING is a seq2seq model for Text-to-AMR and AMR-to-Text (AAAI2021). (139 stars)"
    },
    {
     "name": "maverick-coref",
     "url": "https://github.com/SapienzaNLP/maverick-coref",
     "note": "Open-source repository (71 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "sapienzanlp/Minerva-3B-base-v1.0",
     "url": "https://huggingface.co/sapienzanlp/Minerva-3B-base-v1.0",
     "kind": "model",
     "note": "Model: 5,847 downloads, 58 likes."
    },
    {
     "name": "sapienzanlp/Minerva-350M-base-v1.0",
     "url": "https://huggingface.co/sapienzanlp/Minerva-350M-base-v1.0",
     "kind": "model",
     "note": "Model: 5,601 downloads, 17 likes."
    },
    {
     "name": "sapienzanlp/INDAQA_CALAMITA",
     "url": "https://huggingface.co/datasets/sapienzanlp/INDAQA_CALAMITA",
     "kind": "dataset",
     "note": "Dataset: 804 downloads, 2 likes."
    },
    {
     "name": "sapienzanlp/ea-mt-leaderboard",
     "url": "https://huggingface.co/spaces/sapienzanlp/ea-mt-leaderboard",
     "kind": "space",
     "note": "Space: 4 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "ReLiK: Retrieve and LinK, Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget",
     "url": "https://arxiv.org/abs/2408.00103",
     "year": "2024"
    },
    {
     "title": "Maverick: Efficient and Accurate Coreference Resolution Defying Recent Trends",
     "url": "https://arxiv.org/abs/2407.21489",
     "year": "2024"
    },
    {
     "title": "Minerva LLMs: The First Family of Large Language Models Trained from Scratch on Italian Data",
     "url": "https://aclanthology.org/2024.clicit-1.77.pdf",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Minerva-7B-instruct",
     "url": "https://huggingface.co/sapienzanlp/Minerva-7B-instruct-v1.0",
     "note": "Italian-English instruction model"
    },
    {
     "name": "Minerva-3B-base",
     "url": "https://huggingface.co/sapienzanlp/Minerva-3B-base-v1.0",
     "note": "Base model trained from scratch"
    },
    {
     "name": "ReLiK",
     "url": "https://github.com/SapienzaNLP/relik",
     "note": "Entity linking and relation extraction toolkit"
    }
   ],
   "openness": "Models, datasets and code are public on Hugging Face and GitHub; licences vary per project.",
   "key_people": [
    {
     "name": "Roberto Navigli",
     "role": "Group head, professor",
     "url": "https://www.diag.uniroma1.it/navigli",
     "note": null
    },
    {
     "name": "Edoardo Barba",
     "role": "Minerva project lead",
     "url": "https://nlp.uniroma1.it",
     "note": null
    },
    {
     "name": "Simone Conia",
     "role": "Minerva project lead",
     "url": "https://nlp.uniroma1.it",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://nlp.uniroma1.it",
    "https://huggingface.co/sapienzanlp",
    "https://github.com/SapienzaNLP",
    "https://arxiv.org/abs/2408.00103"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "sapienza-nlp",
   "logo": true
  },
  {
   "orig_id": "sarvam-ai",
   "name": "Sarvam AI",
   "alias": null,
   "full_name": "Sarvam AI",
   "type": "startup",
   "founded": 2023,
   "country": "India",
   "city": "Bengaluru",
   "lat": 12.97,
   "lon": 77.59,
   "website": "https://www.sarvam.ai/",
   "huggingface": "https://huggingface.co/sarvamai",
   "github": "https://github.com/sarvamai",
   "x": "https://twitter.com/SarvamAI",
   "discord": "https://discord.com/invite/5rAsykttcs",
   "linkedin": null,
   "contact_email": null,
   "summary": "Indian startup building models for Indian languages, with several open-weight releases.",
   "tagline": "An Indian AI company building open and sovereign language models for Indian languages, including Sarvam 30B and 105B.",
   "about": [
    "Sarvam AI was founded in July 2023 in Bengaluru by Vivek Raghavan and Pratyush Kumar, both previously involved with AI4Bharat. Raghavan's background includes India's Aadhaar digital identity system, and Kumar led AI4Bharat.",
    "In December 2023 it announced a 41 million dollar Series A led by Lightspeed, with Peak XV Partners and Khosla Ventures. In February 2026 it launched Sarvam 30B and 105B at the India AI Impact Summit, with weights under Apache 2.0.",
    "The company says the models were trained from scratch on 16T and 12T tokens respectively and target Indian languages. Independent benchmarks from Artificial Analysis place the 105B below several peer open models. Sarvam also builds speech and enterprise products."
   ],
   "mission": "Sarvam aims to build full-stack, sovereign AI for India, including open models.",
   "notice": null,
   "known_for": [
    "Sarvam 30B and 105B",
    "Indian-language models",
    "Apache 2.0 open weights",
    "Sovereign AI in India"
   ],
   "focus": [
    "language models",
    "translation",
    "Indic languages"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded in Bengaluru by Vivek Raghavan and Pratyush Kumar.",
     "year": "2023"
    },
    {
     "title": "Series A",
     "note": "41 million dollar round led by Lightspeed.",
     "year": "2023"
    },
    {
     "title": "Sarvam 30B and 105B launched",
     "note": "Introduced at the India AI Impact Summit.",
     "year": "2026"
    },
    {
     "title": "Weights open-sourced",
     "note": "Released under Apache 2.0 on Hugging Face and AI Kosh.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "sarvam-ai-cookbook",
     "url": "https://github.com/sarvamai/sarvam-ai-cookbook",
     "note": "Open-source cookbook for the Sarvam API."
    },
    {
     "name": "skills",
     "url": "https://github.com/sarvamai/skills",
     "note": "Skills and examples for building with Sarvam."
    },
    {
     "name": "llm_intent_entity",
     "url": "https://github.com/sarvamai/llm_intent_entity",
     "note": "Evaluation framework for ASR models."
    },
    {
     "name": "sarvam-mcp",
     "url": "https://github.com/sarvamai/sarvam-mcp",
     "note": "Official Sarvam MCP server."
    }
   ],
   "hf_highlights": [
    {
     "name": "sarvamai/sarvam-translate",
     "url": "https://huggingface.co/sarvamai/sarvam-translate",
     "kind": "model",
     "note": "Most downloaded Sarvam model."
    },
    {
     "name": "sarvamai/sarvam-1",
     "url": "https://huggingface.co/sarvamai/sarvam-1",
     "kind": "model",
     "note": "2B Indic model."
    },
    {
     "name": "sarvamai/sarvam-105b",
     "url": "https://huggingface.co/sarvamai/sarvam-105b",
     "kind": "model",
     "note": "105B open-weight model."
    },
    {
     "name": "sarvamai/mmlu-indic",
     "url": "https://huggingface.co/datasets/sarvamai/mmlu-indic",
     "kind": "dataset",
     "note": "Indic MMLU."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "sarvam-30b",
     "url": "https://huggingface.co/sarvamai/sarvam-30b",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "sarvam-105b",
     "url": "https://huggingface.co/sarvamai/sarvam-105b",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "sarvam-translate",
     "url": "https://huggingface.co/sarvamai/sarvam-translate",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open weights on Hugging Face; no GitHub org listed.",
   "key_people": [
    {
     "name": "Vivek Raghavan",
     "role": "Co-founder",
     "url": "https://www.sarvam.ai/",
     "note": null
    },
    {
     "name": "Pratyush Kumar",
     "role": "Co-founder",
     "url": "https://www.sarvam.ai/",
     "note": null
    }
   ],
   "discord_note": "Developer community Discord linked from sarvam.ai.",
   "contact_note": "sarvam.ai has a Contact Us form and a developer community Discord.",
   "sources": [
    "https://huggingface.co/sarvamai",
    "https://www.sarvam.ai/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "sarvam-ai",
   "logo": true
  },
  {
   "orig_id": "sesame",
   "name": "Sesame",
   "alias": null,
   "full_name": "Sesame",
   "type": "startup",
   "founded": null,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.sesame.com",
   "huggingface": "https://huggingface.co/sesame",
   "github": "https://github.com/SesameAILabs",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Speech AI company that open-sourced the CSM-1B conversational speech model.",
   "tagline": "A conversational speech company that open-sourced the CSM-1B base speech model while building voice companions.",
   "about": [
    "Sesame builds conversational voice AI and is co-led by Brendan Iribe, a co-founder of Oculus, and Ankit Kumar. It announced its voice companions in February 2025 and has since pursued voice-driven wearables, raising 250 million dollars in a round led by Sequoia Capital in October 2025.",
    "In March 2025 it open-sourced CSM-1B, a base speech generation model pairing a Llama backbone with an audio decoder that produces Mimi audio codes. The release is under Apache 2.0 with ethical use guidelines.",
    "The open checkpoint is a base model that has not been fine-tuned to a specific voice; a fine-tuned version powers the company's demo. The technical post lists Johan Schalkwyk, Ankit Kumar, Dan Lyth, Sefik Emre Eskimez, Zack Hodari, Cinjon Resnick, Ramon Sanabria and Raven Jiang."
   ],
   "mission": "Sesame aims for natural, voice-based conversational AI and shares a base model with the community.",
   "notice": null,
   "known_for": [
    "CSM-1B open speech model",
    "Conversational voice AI",
    "Mimi audio codes",
    "Apache 2.0 release"
   ],
   "focus": [
    "speech"
   ],
   "contributions": [
    {
     "title": "Conversational speech announced",
     "note": "Public research preview of a conversational voice model.",
     "year": "2025"
    },
    {
     "title": "CSM-1B open-sourced",
     "note": "Released code and base weights under Apache 2.0.",
     "year": "2025"
    },
    {
     "title": "250 million dollar round",
     "note": "Led by Sequoia Capital, to support voice-driven wearables.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "csm",
     "url": "https://github.com/SesameAILabs/csm",
     "note": "Conversational speech generation model."
    },
    {
     "name": "turnbench",
     "url": "https://github.com/SesameAILabs/turnbench",
     "note": "End-of-turn benchmark."
    },
    {
     "name": "link",
     "url": "https://github.com/SesameAILabs/link",
     "note": "Issue tracker for Sesame releases."
    }
   ],
   "hf_highlights": [
    {
     "name": "csm-1b",
     "url": "https://huggingface.co/sesame/csm-1b",
     "kind": "model",
     "note": "Open 1B speech model."
    },
    {
     "name": "csm-1b",
     "url": "https://huggingface.co/spaces/sesame/csm-1b",
     "kind": "space",
     "note": "Demo space."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "csm",
     "url": "https://github.com/SesameAILabs/csm",
     "note": "CSM code"
    },
    {
     "name": "csm-1b",
     "url": "https://huggingface.co/sesame/csm-1b",
     "note": "Weights"
    }
   ],
   "openness": "Open weights for CSM-1B.",
   "key_people": [
    {
     "name": "Ankit Kumar",
     "role": "Author of CSM release",
     "url": "https://github.com/SesameAILabs/csm",
     "note": null
    },
    {
     "name": "Johan Schalkwyk",
     "role": "Author of CSM release",
     "url": "https://github.com/SesameAILabs/csm",
     "note": null
    },
    {
     "name": "Dan Lyth",
     "role": "Author",
     "url": "https://github.com/SesameAILabs/csm",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://www.sesame.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.sesame.com",
    "https://huggingface.co/sesame"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "sesame",
   "mlat": 37.71869,
   "mlon": -122.55353,
   "logo": true
  },
  {
   "orig_id": "shanghai-ai-lab",
   "name": "Shanghai AI Laboratory",
   "alias": "InternLM",
   "full_name": "Shanghai AI Laboratory (InternLM)",
   "type": "institute",
   "founded": 2020,
   "country": "China",
   "city": "Shanghai",
   "lat": 31.23,
   "lon": 121.47,
   "website": "https://chat.intern-ai.org.cn/",
   "huggingface": "https://huggingface.co/internlm",
   "github": "https://github.com/internlm",
   "x": "https://twitter.com/intern_lm",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Research laboratory that open-sources the Intern family of large language and multimodal models and a full toolchain. Government-backed.",
   "tagline": "A state-backed Chinese research laboratory behind the open Intern family of language and multimodal models: InternLM, InternVL and Intern-S1.",
   "about": [
    "Shanghai AI Laboratory is a non-profit research institution inaugurated at the World AI Conference in Shanghai in 2020. Zhou Bowen became director, appearing in that role in July 2024.",
    "The lab's Intern series includes InternLM multilingual foundation and chat models, InternVL multimodal models, and Intern-S1, a 241B-parameter mixture-of-experts scientific multimodal model with 28B parameters active per token, released under Apache 2.0.",
    "It also open-sources a toolchain around these models."
   ],
   "mission": "The lab aims to advance open general AI, with an emphasis on science.",
   "notice": null,
   "known_for": [
    "InternLM models",
    "InternVL multimodal",
    "Intern-S1 scientific model",
    "Open toolchain"
   ],
   "focus": [
    "language models",
    "multimodal",
    "toolchain",
    "Chinese"
   ],
   "contributions": [
    {
     "title": "Lab inaugurated",
     "note": "Opened at the World AI Conference in Shanghai.",
     "year": "2020"
    },
    {
     "title": "InternLM2 technical report",
     "note": "Described the open InternLM2 base and chat models.",
     "year": "2024"
    },
    {
     "title": "Intern-S1",
     "note": "Released a scientific multimodal foundation model at the 2025 World AI Conference.",
     "year": "2025"
    },
    {
     "title": "Intern-S1-Pro",
     "note": "Released a trillion-parameter mixture-of-experts scientific model in February 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "lmdeploy",
     "url": "https://github.com/InternLM/lmdeploy",
     "note": "Toolkit for compressing, deploying and serving LLMs."
    },
    {
     "name": "InternLM",
     "url": "https://github.com/InternLM/InternLM",
     "note": "Official release of the InternLM series."
    },
    {
     "name": "MindSearch",
     "url": "https://github.com/InternLM/MindSearch",
     "note": "LLM-based multi-agent web search framework."
    },
    {
     "name": "xtuner",
     "url": "https://github.com/InternLM/xtuner",
     "note": "Training engine for large MoE models."
    }
   ],
   "hf_highlights": [
    {
     "name": "internlm-chat-7b",
     "url": "https://huggingface.co/internlm/internlm-chat-7b",
     "kind": "model",
     "note": "Early chat model."
    },
    {
     "name": "internlm2-chat-7b",
     "url": "https://huggingface.co/internlm/internlm2-chat-7b",
     "kind": "model",
     "note": "InternLM2 chat model."
    },
    {
     "name": "internlm3-8b-instruct",
     "url": "https://huggingface.co/internlm/internlm3-8b-instruct",
     "kind": "model",
     "note": "InternLM3 instruct model."
    },
    {
     "name": "WildClawBench",
     "url": "https://huggingface.co/datasets/internlm/WildClawBench",
     "kind": "dataset",
     "note": "Agent benchmark."
    }
   ],
   "papers": [
    {
     "title": "Intern-S1: A Scientific Multimodal Foundation Model",
     "url": "https://arxiv.org/abs/2508.15763",
     "year": "2025"
    },
    {
     "title": "InternLM2 Technical Report",
     "url": "https://arxiv.org/abs/2403.17297",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Sage-S2.1-Flash",
     "url": "https://huggingface.co/internlm/Sage-S2.1-Flash",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "InternLumina-U2",
     "url": "https://huggingface.co/internlm/InternLumina-U2",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Open weights on Hugging Face and open code on GitHub.",
   "key_people": [
    {
     "name": "Zhou Bowen",
     "role": "Director",
     "url": "https://web.ee.tsinghua.edu.cn/zhoubowen/en/kyxm/8232/content/1717.htm",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website or GitHub and Hugging Face pages.",
   "sources": [
    "https://huggingface.co/internlm",
    "https://chat.intern-ai.org.cn/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "shanghai-ai-laboratory",
   "mlat": 31.28,
   "mlon": 121.47,
   "logo": true
  },
  {
   "orig_id": "softcatala",
   "name": "Softcatala",
   "alias": null,
   "full_name": "Softcatala",
   "type": "nonprofit",
   "founded": 1998,
   "country": "Spain",
   "city": "Barcelona",
   "lat": 41.3874,
   "lon": 2.1686,
   "website": "https://www.softcatala.org",
   "huggingface": "https://huggingface.co/softcatala",
   "github": "https://github.com/Softcatala",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Catalan non-profit promoting Catalan in technology with open machine translation, speech and dictionary tools.",
   "tagline": "A volunteer-driven Catalan non-profit that makes software, translation and speech tools available in Catalan.",
   "about": [
    "Softcatala was created in 1998 by volunteers to promote Catalan in computing and on the Internet. Jordi Mas is named as its founder. Early work focused on translating proprietary software before shifting toward free software.",
    "It operates a machine translation service, available since 2000, which moved from a non-free engine to Apertium. The group helps maintain Apertium language pairs by adding vocabulary and rules.",
    "It publishes models and datasets on Hugging Face and code on GitHub, and has a long record of localised versions of popular programs."
   ],
   "mission": "The association aims to ensure Catalan speakers can use technology fully in their own language.",
   "notice": null,
   "known_for": [
    "Catalan localisation of free software",
    "Catalan machine translation service",
    "Apertium contributions"
   ],
   "focus": [
    "Catalan",
    "machine translation",
    "speech"
   ],
   "contributions": [
    {
     "title": "Softcatala founded",
     "note": "Volunteers began promoting Catalan in computing.",
     "year": "1998"
    },
    {
     "title": "Machine translation service",
     "note": "Launched online Catalan translation.",
     "year": "2000"
    },
    {
     "title": "Move to Apertium",
     "note": "Switched the translator to the free Apertium engine.",
     "year": "2011"
    },
    {
     "title": "Catalan Whisper models",
     "note": "Fine-tuned whisper-base and whisper-small for Catalan during the Hugging Face Whisper sprint, under Apache 2.0.",
     "year": "2023"
    },
    {
     "title": "25th anniversary",
     "note": "Marked 25 years of work.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "whisper-ctranslate2",
     "url": "https://github.com/Softcatala/whisper-ctranslate2",
     "note": "Whisper command line client compatible with original OpenAI client based on CTranslate2. (1,375 stars)"
    },
    {
     "name": "catalan-dict-tools",
     "url": "https://github.com/Softcatala/catalan-dict-tools",
     "note": "Tools for managing Catalan dictionaries (65 stars)"
    },
    {
     "name": "nmt-softcatala",
     "url": "https://github.com/Softcatala/nmt-softcatala",
     "note": "This repository contains Neural Machine Translation tools built at Softcatalà  (45 stars)"
    },
    {
     "name": "LanguageToolAndroidService",
     "url": "https://github.com/Softcatala/LanguageToolAndroidService",
     "note": "Experiments with spell checking on Android (33 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "softcatala/wav2vec2-large-xlsr-catala",
     "url": "https://huggingface.co/softcatala/wav2vec2-large-xlsr-catala",
     "kind": "model",
     "note": "Model with 4 likes on the Hub."
    },
    {
     "name": "softcatala/translate-cat-eng",
     "url": "https://huggingface.co/softcatala/translate-cat-eng",
     "kind": "model",
     "note": "Model with 3 likes on the Hub."
    },
    {
     "name": "softcatala/catalan-dictionary",
     "url": "https://huggingface.co/datasets/softcatala/catalan-dictionary",
     "kind": "dataset",
     "note": "Dataset with 3 likes on the Hub."
    },
    {
     "name": "softcatala/comparativa-tts-catala",
     "url": "https://huggingface.co/spaces/softcatala/comparativa-tts-catala",
     "kind": "space",
     "note": "Space with 4 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "whisper-ctranslate2",
     "url": "https://github.com/Softcatala/whisper-ctranslate2",
     "note": "Whisper CLI based on CTranslate2"
    },
    {
     "name": "nmt-softcatala",
     "url": "https://github.com/Softcatala/nmt-softcatala",
     "note": "Neural machine translation tools"
    },
    {
     "name": "Catalan Whisper",
     "url": "https://huggingface.co/softcatala/whisper-medium-ca",
     "note": "Whisper for Catalan"
    },
    {
     "name": "Catalan speech",
     "url": "https://huggingface.co/datasets/softcatala/catalan-youtube-speech",
     "note": "Catalan speech dataset"
    }
   ],
   "openness": "Open-source tools, models and datasets on GitHub and Hugging Face.",
   "key_people": [
    {
     "name": "Jordi Mas",
     "role": "Founder",
     "url": "https://en.wikipedia.org/wiki/Softcatal%C3%A0",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the organisation website.",
   "sources": [
    "https://www.softcatala.org",
    "https://huggingface.co/softcatala",
    "https://github.com/Softcatala"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "softcatala",
   "mlat": 41.3874,
   "mlon": 2.10196,
   "logo": true
  },
  {
   "orig_id": "speakleash",
   "name": "SpeakLeash",
   "alias": "Spichlerz",
   "full_name": "SpeakLeash (Spichlerz)",
   "type": "collective",
   "founded": 2022,
   "country": "Poland",
   "city": "Warsaw",
   "lat": 52.2297,
   "lon": 21.0122,
   "website": "https://speakleash.org",
   "huggingface": "https://huggingface.co/speakleash",
   "github": "https://github.com/speakleash",
   "x": "https://x.com/Speak_Leash",
   "discord": "https://discord.gg/ZJwCMrxwT7",
   "linkedin": "https://www.linkedin.com/company/speakleash",
   "contact_email": null,
   "summary": "Open community project collecting Polish text data and building Polish language models. It is the group behind the Bielik models.",
   "tagline": "A Polish open community collecting Polish text and building the Bielik models with ACK Cyfronet AGH.",
   "about": [
    "SpeakLeash, also known as Spichlerz, is an open community project founded in 2022 that collects Polish text for language modeling. Sebastian Kondracki is described as co-creator.",
    "The Bielik family of Polish models, from 1.5B to 11B parameters, was developed with ACK Cyfronet AGH, which provides the Helios supercomputer. Bielik 11B v2 has a technical report on arXiv, and Bielik 11B v3 followed in 2025 under Apache 2.0, with authors including Krzysztof Ociepa, Łukasz Flis, Remigiusz Kinas, Krzysztof Wróbel and Adrian Gwoździej.",
    "Models and datasets are published on Hugging Face."
   ],
   "mission": "SpeakLeash aims to give Polish a strong open language model ecosystem.",
   "notice": null,
   "known_for": [
    "Bielik Polish models",
    "Polish text corpus",
    "Cyfronet collaboration",
    "Open community data"
   ],
   "focus": [
    "language models",
    "Polish",
    "datasets",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "GitHub organization created",
     "note": "Community data collection begins.",
     "year": "2022"
    },
    {
     "title": "Bielik 7B v0.1",
     "note": "First Bielik model with a technical report.",
     "year": "2024"
    },
    {
     "title": "Bielik 11B v2",
     "note": "Technical report published.",
     "year": "2025"
    },
    {
     "title": "Bielik 11B v3",
     "note": "Instruct model released under Apache 2.0 on Hugging Face.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "Bielik-how-to-start",
     "url": "https://github.com/speakleash/Bielik-how-to-start",
     "note": "Skrypty, tutoriale oraz programistyczna baza wiedzy dotycząca pracy z modelem Bielik. (199 stars)"
    },
    {
     "name": "speakleash",
     "url": "https://github.com/speakleash/speakleash",
     "note": "Open-source repository (84 stars)"
    },
    {
     "name": "bielik-tools",
     "url": "https://github.com/speakleash/bielik-tools",
     "note": "Open-source repository (46 stars)"
    },
    {
     "name": "eskadra-bielik-misja1",
     "url": "https://github.com/speakleash/eskadra-bielik-misja1",
     "note": "Eskadra Bielika - Misja 1 - repozytorium główne (44 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "speakleash/Bielik-11B-v3.0-Instruct",
     "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct",
     "kind": "model",
     "note": "Model: 103,623 downloads, 160 likes."
    },
    {
     "name": "speakleash/Bielik-11B-v2.3-Instruct",
     "url": "https://huggingface.co/speakleash/Bielik-11B-v2.3-Instruct",
     "kind": "model",
     "note": "Model: 10,877 downloads, 52 likes."
    },
    {
     "name": "speakleash/PES-2018-2022",
     "url": "https://huggingface.co/datasets/speakleash/PES-2018-2022",
     "kind": "dataset",
     "note": "Dataset: 226 downloads, 1 likes."
    },
    {
     "name": "speakleash/open_pl_llm_leaderboard",
     "url": "https://huggingface.co/spaces/speakleash/open_pl_llm_leaderboard",
     "kind": "space",
     "note": "Space: 85 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Bielik 7B v0.1: A Polish Language Model -- Development, Insights, and Evaluation",
     "url": "https://arxiv.org/abs/2410.18565",
     "year": "2024"
    },
    {
     "title": "Bielik v3 Small: Technical Report",
     "url": "https://arxiv.org/abs/2505.02550",
     "year": "2025"
    },
    {
     "title": "Bielik 11B v2 Technical Report",
     "url": "https://arxiv.org/abs/2505.02410",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Bielik-11B-v3.0-Instruct",
     "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct",
     "note": "Polish instruction model"
    },
    {
     "name": "Bielik-7B-v0.1",
     "url": "https://huggingface.co/speakleash/Bielik-7B-v0.1",
     "note": "First Bielik base model"
    },
    {
     "name": "Open PL LLM Leaderboard",
     "url": "https://huggingface.co/spaces/speakleash/open_pl_llm_leaderboard",
     "note": "Leaderboard for Polish LLMs"
    }
   ],
   "openness": "Bielik models and tools are public on Hugging Face and GitHub, with MIT-licensed code.",
   "key_people": [
    {
     "name": "Sebastian Kondracki",
     "role": "Co-creator",
     "url": "https://speakleash.org",
     "note": null
    },
    {
     "name": "Marek Magryś",
     "role": "Deputy director at Cyfronet, supporting Bielik",
     "url": "https://www.cyfronet.pl/en/projects/initiatives/llm",
     "note": null
    }
   ],
   "discord_note": "Community Discord of the SpeakLeash project, for Bielik users and data contributors.",
   "contact_note": "Discord and website forms; no email found on the homepage.",
   "sources": [
    "https://speakleash.org",
    "https://huggingface.co/speakleash",
    "https://github.com/speakleash",
    "https://arxiv.org/abs/2410.18565"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "speakleash",
   "mlat": 52.26506,
   "mlon": 21.0122,
   "logo": true
  },
  {
   "orig_id": "spraakbanken",
   "name": "Sprakbanken Text",
   "alias": "University of Gothenburg",
   "full_name": "Sprakbanken Text (University of Gothenburg)",
   "type": "academic",
   "founded": 1975,
   "country": "Sweden",
   "city": "Gothenburg",
   "lat": 57.7089,
   "lon": 11.9746,
   "website": "https://spraakbanken.gu.se",
   "huggingface": null,
   "github": "https://github.com/spraakbanken",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Sprakbanken Text at the University of Gothenburg develops Swedish language resources and open tools such as the Sparv text analysis pipeline and the Korp corpus search.",
   "tagline": "The University of Gothenburg's national language data infrastructure, building Korp, Sparv and open resources for Swedish since 1975.",
   "about": [
    "Språkbanken Text is a research and development unit at the University of Gothenburg, established in 1975 and among the world's oldest language technology units. It is now being renamed simply Språkbanken within a national infrastructure.",
    "It builds research platforms including Korp (corpus search), Karp (lexicons), Mink, Sparv (text analysis) and Strix, and focuses on Swedish across historical periods. Lexical resources are released under CC-BY.",
    "It takes part in CLARIN and HumInfra, and the national infrastructure is moving its news and data services to sprakbanken.se."
   ],
   "mission": "The unit develops and shares language data and tools so researchers and society can work with Swedish text.",
   "notice": null,
   "known_for": [
    "Korp corpus search",
    "Sparv annotation pipeline",
    "Oldest language technology unit"
   ],
   "focus": [
    "Swedish NLP",
    "corpora"
   ],
   "contributions": [
    {
     "title": "Sprakbanken established",
     "note": "Government mandate to University of Gothenburg.",
     "year": "1975"
    },
    {
     "title": "Lexicography split",
     "note": "Lexicographical work moved to a separate centre around the turn of the millennium.",
     "year": "2000"
    },
    {
     "title": "Korp",
     "note": "Corpus search covers about 15 billion words of Swedish.",
     "year": "2020"
    },
    {
     "title": "50th anniversary programme",
     "note": "Sprakbanken marked 50 years.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "sparv",
     "url": "https://github.com/spraakbanken/sparv",
     "note": "Språkbanken's text analysis tool (29 stars)"
    },
    {
     "name": "korp-backend",
     "url": "https://github.com/spraakbanken/korp-backend",
     "note": "Backend for Korp, Språkbanken's corpus search tool (17 stars)"
    },
    {
     "name": "korp-frontend",
     "url": "https://github.com/spraakbanken/korp-frontend",
     "note": "Frontend for Korp, a tool using the IMS Open Corpus Workbench (CWB). (16 stars)"
    },
    {
     "name": "multiged-2023",
     "url": "https://github.com/spraakbanken/multiged-2023",
     "note": "Open-source repository. (15 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Sparv",
     "url": "https://github.com/spraakbanken/sparv",
     "note": "Text analysis pipeline"
    },
    {
     "name": "Korp backend",
     "url": "https://github.com/spraakbanken/korp-backend",
     "note": "Backend for the Korp corpus search tool"
    },
    {
     "name": "Korp frontend",
     "url": "https://github.com/spraakbanken/korp-frontend",
     "note": "Web frontend for corpus search"
    },
    {
     "name": "multiged-2023",
     "url": "https://github.com/spraakbanken/multiged-2023",
     "note": "Multilingual grammatical error detection shared task resources"
    }
   ],
   "openness": "Tools are open source on GitHub; corpora are distributed through the Sprakbanken service.",
   "key_people": [
    {
     "name": "Lars Borin",
     "role": "Professor of NLP",
     "url": "https://spraakbanken.gu.se/en/about/staff",
     "note": null
    },
    {
     "name": "Markus Forsberg",
     "role": "Director of the department and national director of Språkbanken",
     "url": "https://www.gu.se/en/about/find-staff/markusforsberg",
     "note": null
    },
    {
     "name": "Aleksandrs Berdicevskis",
     "role": "Vice director and associate professor",
     "url": "https://www.gu.se/en/about/find-staff/aleksandrsberdicevskis",
     "note": null
    },
    {
     "name": "Dana Dannélls",
     "role": "Researcher and vice director",
     "url": "https://www.gu.se/en/about/find-staff/danadannells",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://spraakbanken.gu.se",
    "https://github.com/spraakbanken"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "sprakbanken-text",
   "mlat": 57.74426,
   "mlon": 11.9746,
   "logo": true
  },
  {
   "orig_id": "eth-sri",
   "name": "SRI Lab",
   "alias": "ETH Zurich",
   "full_name": "SRI Lab (ETH Zurich)",
   "type": "academic",
   "founded": null,
   "country": "Switzerland",
   "city": "Zurich",
   "lat": 47.3769,
   "lon": 8.5417,
   "website": "https://www.sri.inf.ethz.ch",
   "huggingface": null,
   "github": "https://github.com/eth-sri",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Secure, Reliable and Intelligent Systems Lab at ETH Zurich, publishing open code for LLM programming (LMQL), neural network verification and ML for code.",
   "tagline": "ETH Zurich's group on trustworthy machine learning, producing neural network verifiers, LMQL and ML-for-code tools.",
   "about": [
    "The Secure, Reliable and Intelligent Systems Lab at ETH Zurich is led by Martin Vechev. It works on program analysis, verification and the reliability of machine learning.",
    "Its AI2 paper at IEEE S&P 2018 introduced abstract interpretation for certifying neural networks, and the ETH Robustness Analyzer (ERAN) followed.",
    "Its code is published on GitHub, including LMQL, a language for programming with language models."
   ],
   "mission": "The lab works to make intelligent systems safe, secure and reliable by combining machine learning with program analysis.",
   "notice": null,
   "known_for": [
    "AI2 neural network certification",
    "ERAN robustness analyzer",
    "LMQL query language"
   ],
   "focus": [
    "LLM programming",
    "robustness"
   ],
   "contributions": [
    {
     "title": "AI2",
     "note": "Safety and robustness certification with abstract interpretation at IEEE S&P.",
     "year": "2018"
    },
    {
     "title": "ERAN released",
     "note": "Robustness analyzer for deep neural networks.",
     "year": "2019"
    },
    {
     "title": "DeepCode and ChainSecurity acquired",
     "note": "Vechev's spinouts DeepCode (by Snyk) and ChainSecurity (by PwC) were acquired; LatticeFlow was co-founded the same year.",
     "year": "2020"
    },
    {
     "title": "LMQL",
     "note": "Language-model programming language published.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "lmql",
     "url": "https://github.com/eth-sri/lmql",
     "note": "A language for constraint-guided and efficient LLM programming. (4,228 stars)"
    },
    {
     "name": "securify2",
     "url": "https://github.com/eth-sri/securify2",
     "note": "Securify v2.0 (636 stars)"
    },
    {
     "name": "debin",
     "url": "https://github.com/eth-sri/debin",
     "note": "Machine Learning to Deobfuscate Binaries (453 stars)"
    },
    {
     "name": "eran",
     "url": "https://github.com/eth-sri/eran",
     "note": "ETH Robustness Analyzer for Deep Neural Networks (348 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "Prompting Is Programming: A Query Language for Large Language Models",
     "url": "https://arxiv.org/abs/2212.06094",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "LMQL",
     "url": "https://github.com/eth-sri/lmql",
     "note": "Query language for LLM programming (4,000+ stars)"
    },
    {
     "name": "ERAN",
     "url": "https://github.com/eth-sri/eran",
     "note": "Robustness analyzer for deep networks"
    },
    {
     "name": "Debin",
     "url": "https://github.com/eth-sri/debin",
     "note": "ML to deobfuscate binaries"
    },
    {
     "name": "Securify2",
     "url": "https://github.com/eth-sri/securify2",
     "note": "Smart contract security analyzer"
    }
   ],
   "openness": "Tools are public on GitHub, mostly under open-source licences.",
   "key_people": [
    {
     "name": "Martin Vechev",
     "role": "Professor and head of the lab",
     "url": "https://www.sri.inf.ethz.ch/people/martin",
     "note": "Founder of DeepCode, ChainSecurity and LatticeFlow; architect of INSAIT."
    },
    {
     "name": "Timon Gehr",
     "role": "Senior scientist",
     "url": "https://www.sri.inf.ethz.ch/people/timon",
     "note": "Senior scientist at the lab."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.sri.inf.ethz.ch",
    "https://github.com/eth-sri",
    "https://arxiv.org/abs/2212.06094"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "sri-lab",
   "mlat": 47.39855,
   "mlon": 8.59708,
   "logo": true
  },
  {
   "orig_id": "stability-ai",
   "name": "Stability AI",
   "alias": null,
   "full_name": "Stability AI",
   "type": "startup",
   "founded": 2019,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://stability.ai",
   "huggingface": "https://huggingface.co/stabilityai",
   "github": "https://github.com/Stability-AI",
   "x": "https://x.com/StabilityAI",
   "discord": "https://discord.gg/stablediffusion",
   "linkedin": "https://www.linkedin.com/company/stability-ai",
   "contact_email": "press@stability.ai",
   "summary": "London-based generative AI company known for Stable Diffusion. It publishes open-weight image, video, audio and language models, many under community licences.",
   "tagline": "A London generative AI company behind Stable Diffusion that publishes open-weight image, video, audio and language models, many under community licences.",
   "about": [
    "Stability AI was founded by Emad Mostaque, who served as CEO until March 2024. Prem Akkaraju became CEO in June 2024, with Sean Parker as executive chairman, after a restructuring.",
    "Stable Diffusion 1.5 (October 2022) became a foundation for community fine-tuning. Stable Diffusion 3.5 followed in late 2024 under a community licence free for those with under $1M annual revenue.",
    "Not all releases are permissively licensed."
   ],
   "mission": "Stability aims to provide generative models across modalities with open weights.",
   "notice": null,
   "known_for": [
    "Stable Diffusion",
    "Open-weight image models",
    "Stable Diffusion 3.5",
    "Community licence"
   ],
   "focus": [
    "image generation",
    "video",
    "audio",
    "language models"
   ],
   "contributions": [
    {
     "title": "Stable Diffusion public release",
     "note": "Open weights for a latent diffusion text-to-image model, developed with CompVis and Runway.",
     "year": "2022"
    },
    {
     "title": "SDXL",
     "note": "Larger latent diffusion model released with open weights.",
     "year": "2023"
    },
    {
     "title": "Stable Video Diffusion",
     "note": "Open-weight image-to-video model.",
     "year": "2023"
    },
    {
     "title": "Stable Diffusion 3 Medium",
     "note": "Open weights under a community licence.",
     "year": "2024"
    },
    {
     "title": "Stable Diffusion 3.5",
     "note": "Large and Medium variants under the Stability Community License.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "generative-models",
     "url": "https://github.com/Stability-AI/generative-models",
     "note": "Generative Models by Stability AI (27,344 stars)"
    },
    {
     "name": "StableLM",
     "url": "https://github.com/Stability-AI/StableLM",
     "note": "StableLM: Stability AI Language Models (15,644 stars)"
    },
    {
     "name": "StableStudio",
     "url": "https://github.com/Stability-AI/StableStudio",
     "note": "Community interface for generative AI (9,033 stars)"
    },
    {
     "name": "StableCascade",
     "url": "https://github.com/Stability-AI/StableCascade",
     "note": "Official Code for Stable Cascade (6,518 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "stabilityai/stable-diffusion-xl-base-1.0",
     "url": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0",
     "kind": "model",
     "note": "Model: 4,797,783 downloads, 8480 likes."
    },
    {
     "name": "stabilityai/sdxl-turbo",
     "url": "https://huggingface.co/stabilityai/sdxl-turbo",
     "kind": "model",
     "note": "Model: 583,716 downloads, 2673 likes."
    },
    {
     "name": "stabilityai/stable-diffusion-3.5-large",
     "url": "https://huggingface.co/spaces/stabilityai/stable-diffusion-3.5-large",
     "kind": "space",
     "note": "Space: 2053 likes on the Hub."
    },
    {
     "name": "stabilityai/sd-turbo",
     "url": "https://huggingface.co/stabilityai/sd-turbo",
     "kind": "model",
     "note": "Model: 389,992 downloads, 471 likes."
    }
   ],
   "papers": [
    {
     "title": "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis",
     "url": "https://arxiv.org/abs/2403.03206",
     "year": "2024"
    },
    {
     "title": "SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis",
     "url": "https://arxiv.org/abs/2307.01952",
     "year": "2023"
    },
    {
     "title": "Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets",
     "url": "https://arxiv.org/abs/2311.15127",
     "year": "2023"
    },
    {
     "title": "Stable LM 2 1.6B Technical Report",
     "url": "https://arxiv.org/abs/2402.17834",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Stable Diffusion XL",
     "url": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0",
     "note": "Open-weight text-to-image model"
    },
    {
     "name": "Stable Diffusion 3.5 Large",
     "url": "https://huggingface.co/stabilityai/stable-diffusion-3.5-large",
     "note": "Latest open-weight diffusion transformer family"
    },
    {
     "name": "StableLM",
     "url": "https://github.com/Stability-AI/StableLM",
     "note": "Open language models"
    }
   ],
   "openness": "Weights are public on Hugging Face under a mix of permissive and Stability community licences; code is mostly MIT or Apache-2.0.",
   "key_people": [
    {
     "name": "Emad Mostaque",
     "role": "Founder, CEO until March 2024",
     "url": "https://en.wikipedia.org/wiki/Emad_Mostaque",
     "note": "Led the company through the Stable Diffusion releases."
    },
    {
     "name": "Prem Akkaraju",
     "role": "CEO since June 2024",
     "url": "https://stability.ai/news/category/Company",
     "note": "Took over after the restructuring."
    },
    {
     "name": "Sean Parker",
     "role": "Executive chairman",
     "url": "https://en.wikipedia.org/wiki/Sean_Parker",
     "note": "Joined the board in 2024."
    }
   ],
   "discord_note": "Community Discord for Stable Diffusion users and creators.",
   "contact_note": "Press and partner addresses printed on the website; Discord community.",
   "sources": [
    "https://stability.ai",
    "https://huggingface.co/stabilityai",
    "https://github.com/Stability-AI",
    "https://arxiv.org/abs/2403.03206"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "stability-ai",
   "mlat": 51.52212,
   "mlon": -0.23141,
   "logo": true
  },
  {
   "orig_id": "stanford-crfm",
   "name": "Stanford CRFM",
   "alias": null,
   "full_name": "Stanford CRFM",
   "type": "academic",
   "founded": 2021,
   "country": "United States",
   "city": "Stanford",
   "lat": 37.4275,
   "lon": -122.1697,
   "website": "https://crfm.stanford.edu",
   "huggingface": "https://huggingface.co/stanford-crfm",
   "github": "https://github.com/stanford-crfm",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Stanford center studying foundation models; releases HELM evaluation, BioMedLM and Mistral training code.",
   "tagline": "Stanford's Center for Research on Foundation Models, creators of the HELM evaluation framework and Alpaca.",
   "about": [
    "CRFM is a Stanford HAI center directed by Percy Liang. It studies foundation models and builds open tools for evaluating them.",
    "HELM, published in 2022, benchmarked 30 models on 42 scenarios and 7 metrics under standard conditions, and is maintained as a living benchmark. The center also produced Alpaca, a fine-tuned LLaMA 7B model.",
    "It releases BioMedLM and training code on GitHub and Hugging Face."
   ],
   "mission": "CRFM aims to make foundation models more transparent and understood through open evaluation and research.",
   "notice": null,
   "known_for": [
    "HELM evaluation",
    "Alpaca",
    "BioMedLM",
    "Foundation model research"
   ],
   "focus": [
    "evaluation",
    "language models",
    "policy"
   ],
   "contributions": [
    {
     "title": "CRFM founded",
     "note": "Stanford HAI center for research on foundation models established.",
     "year": "2021"
    },
    {
     "title": "On the Opportunities and Risks of Foundation Models",
     "note": "Report coining the term and surveying the field.",
     "year": "2021"
    },
    {
     "title": "HELM",
     "note": "Published Holistic Evaluation of Language Models.",
     "year": "2022"
    },
    {
     "title": "Alpaca",
     "note": "Released an instruction-tuned LLaMA 7B model.",
     "year": "2023"
    },
    {
     "title": "Foundation Model Transparency Index",
     "note": "Scored major developers on disclosure.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "helm",
     "url": "https://github.com/stanford-crfm/helm",
     "note": "Holistic Evaluation of Language Models."
    },
    {
     "name": "BioMedLM",
     "url": "https://github.com/stanford-crfm/BioMedLM",
     "note": "Biomedical LM."
    },
    {
     "name": "mistral",
     "url": "https://github.com/stanford-crfm/mistral",
     "note": "Framework for transparent GPT-2 training."
    },
    {
     "name": "ecosystem-graphs",
     "url": "https://github.com/stanford-crfm/ecosystem-graphs",
     "note": "Foundation model ecosystem data."
    }
   ],
   "hf_highlights": [
    {
     "name": "air-bench-2024",
     "url": "https://huggingface.co/datasets/stanford-crfm/air-bench-2024",
     "kind": "dataset",
     "note": "Safety benchmark."
    },
    {
     "name": "image2struct-latex-v1",
     "url": "https://huggingface.co/datasets/stanford-crfm/image2struct-latex-v1",
     "kind": "dataset",
     "note": "Image-to-structure benchmark."
    },
    {
     "name": "music-small-800k",
     "url": "https://huggingface.co/stanford-crfm/music-small-800k",
     "kind": "model",
     "note": "Music model."
    }
   ],
   "papers": [
    {
     "title": "Holistic Evaluation of Language Models",
     "url": "https://arxiv.org/abs/2211.09110",
     "year": "2022"
    },
    {
     "title": "On the Opportunities and Risks of Foundation Models",
     "url": "https://arxiv.org/abs/2108.07258",
     "year": "2021"
    },
    {
     "title": "The Foundation Model Transparency Index",
     "url": "https://arxiv.org/abs/2310.12941",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "HELM",
     "url": "https://github.com/stanford-crfm/helm",
     "note": "Evaluation framework"
    },
    {
     "name": "BioMedLM",
     "url": "https://huggingface.co/stanford-crfm/BioMedLM",
     "note": "Biomedical LM"
    }
   ],
   "openness": "Open code and models.",
   "key_people": [
    {
     "name": "Percy Liang",
     "role": "Director",
     "url": "https://cs.stanford.edu/~pliang/",
     "note": "Associate professor of computer science; leads HELM."
    },
    {
     "name": "Rishi Bommasani",
     "role": "Society lead and researcher",
     "url": "https://rishibommasani.github.io/",
     "note": "Lead author of the foundation models report and the Foundation Model Transparency Index."
    },
    {
     "name": "Tatsunori Hashimoto",
     "role": "Faculty researcher",
     "url": "https://thashim.github.io/",
     "note": "Co-led the Alpaca project."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://crfm.stanford.edu); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://crfm.stanford.edu",
    "https://github.com/stanford-crfm"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "stanford-crfm",
   "mlat": 37.4775,
   "mlon": -122.1697,
   "logo": true
  },
  {
   "orig_id": "stanford-nlp",
   "name": "Stanford NLP Group",
   "alias": null,
   "full_name": "Stanford NLP Group",
   "type": "academic",
   "founded": 2000,
   "country": "United States",
   "city": "Stanford",
   "lat": 37.4275,
   "lon": -122.1697,
   "website": "https://nlp.stanford.edu",
   "huggingface": "https://huggingface.co/stanfordnlp",
   "github": "https://github.com/stanfordnlp",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Academic research group releasing widely used open NLP software and datasets, including DSPy, CoreNLP, Stanza and GloVe.",
   "tagline": "The Stanford NLP Group is a leading academic research group whose open software, including CoreNLP, GloVe, Stanza and DSPy, underpins much of modern language technology.",
   "about": [
    "The Stanford NLP Group was started in 2000 by Christopher Manning, who held a joint appointment in linguistics and computer science. It grew to include faculty such as Dan Jurafsky, who arrived in 2004, and Percy Liang, who joined in 2012. The group is part of Stanford University and brings together linguists and computer scientists.",
    "Its software history begins with the Stanford Parser, released as open source in December 2002, and the first unified CoreNLP package appeared on 1 November 2010. Later releases include GloVe word vectors, the Python library Stanza for many human languages, and the DSPy framework for programming rather than prompting language models.",
    "The group publishes code, data and papers openly, and its GitHub organization is among the most widely used sources of NLP tooling for researchers, students and industry engineers."
   ],
   "mission": "To advance the computational understanding of human language and to share research software and datasets that others can build on.",
   "notice": null,
   "known_for": [
    "CoreNLP Java toolkit",
    "GloVe word vectors",
    "Stanza multilingual NLP library",
    "DSPy framework for LM programs"
   ],
   "focus": [
    "NLP",
    "LM programming",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Stanford NLP Group formed",
     "note": "Christopher Manning starts the group at Stanford.",
     "year": "2000"
    },
    {
     "title": "Stanford Parser open-sourced",
     "note": "The parser is first released as open source in December 2002.",
     "year": "2002"
    },
    {
     "title": "CoreNLP unified release",
     "note": "The first unified CoreNLP package appears on 1 November 2010.",
     "year": "2010"
    },
    {
     "title": "GloVe released",
     "note": "Software and data for GloVe word representations are published.",
     "year": "2014"
    },
    {
     "title": "Stanza",
     "note": "A neural Python NLP library covering many human languages is released.",
     "year": "2020"
    },
    {
     "title": "DSPy",
     "note": "The framework for programming language models becomes a flagship open-source project.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "dspy",
     "url": "https://github.com/stanfordnlp/dspy",
     "note": "Programming language models"
    },
    {
     "name": "CoreNLP",
     "url": "https://github.com/stanfordnlp/CoreNLP",
     "note": "Java NLP suite"
    },
    {
     "name": "stanza",
     "url": "https://github.com/stanfordnlp/stanza",
     "note": "Python NLP toolkit"
    },
    {
     "name": "GloVe",
     "url": "https://github.com/stanfordnlp/GloVe",
     "note": "Global vectors for word representation"
    }
   ],
   "hf_highlights": [
    {
     "name": "stanza-en",
     "url": "https://huggingface.co/stanfordnlp/stanza-en",
     "kind": "model",
     "note": "Stanza English"
    },
    {
     "name": "imdb",
     "url": "https://huggingface.co/datasets/stanfordnlp/imdb",
     "kind": "dataset",
     "note": "IMDB sentiment dataset"
    },
    {
     "name": "sst2",
     "url": "https://huggingface.co/datasets/stanfordnlp/sst2",
     "kind": "dataset",
     "note": "SST-2"
    },
    {
     "name": "SHP",
     "url": "https://huggingface.co/datasets/stanfordnlp/SHP",
     "kind": "dataset",
     "note": "Stanford Human Preferences"
    }
   ],
   "papers": [
    {
     "title": "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines",
     "url": "https://arxiv.org/abs/2310.03714",
     "year": "2023"
    },
    {
     "title": "Stanza: A Python Natural Language Processing Toolkit for Many Human Languages",
     "url": "https://arxiv.org/abs/2003.07082",
     "year": "2020"
    }
   ],
   "notable_releases": [
    {
     "name": "dspy",
     "url": "https://github.com/stanfordnlp/dspy",
     "note": "Programming language models"
    },
    {
     "name": "CoreNLP",
     "url": "https://github.com/stanfordnlp/CoreNLP",
     "note": "Java NLP suite"
    },
    {
     "name": "stanza",
     "url": "https://github.com/stanfordnlp/stanza",
     "note": "Python NLP toolkit"
    }
   ],
   "openness": "Mostly MIT or Apache-2.0 code; many datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Christopher Manning",
     "role": "Founder of the group; Professor of Linguistics and Computer Science",
     "url": "https://nlp.stanford.edu/~manning/",
     "note": null
    },
    {
     "name": "Dan Jurafsky",
     "role": "Professor of Linguistics and Computer Science",
     "url": "https://web.stanford.edu/~jurafsky/",
     "note": null
    },
    {
     "name": "Percy Liang",
     "role": "Associate Professor of Computer Science",
     "url": "https://cs.stanford.edu/~pliang/",
     "note": null
    },
    {
     "name": "Christopher Potts",
     "role": "Professor of Linguistics",
     "url": "https://web.stanford.edu/~cgpotts/",
     "note": null
    },
    {
     "name": "Tatsunori Hashimoto",
     "role": "Assistant Professor of Computer Science",
     "url": "https://thashim.github.io/",
     "note": null
    },
    {
     "name": "Diyi Yang",
     "role": "Assistant Professor of Computer Science",
     "url": "https://cs.stanford.edu/~diyiy/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Via website.",
   "sources": [
    "https://github.com/stanfordnlp",
    "https://huggingface.co/stanfordnlp"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "stanford-nlp-group",
   "mlat": 37.4275,
   "mlon": -122.23266,
   "logo": true
  },
  {
   "orig_id": "stepfun",
   "name": "StepFun",
   "alias": null,
   "full_name": "StepFun",
   "type": "startup",
   "founded": 2023,
   "country": "China",
   "city": "Shanghai",
   "lat": 31.23,
   "lon": 121.47,
   "website": "https://www.stepfun.com/",
   "huggingface": "https://huggingface.co/stepfun-ai",
   "github": "https://github.com/stepfun-ai",
   "x": "https://twitter.com/StepFun_ai",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Chinese foundation-model company with 50+ models and a dozen datasets on Hugging Face.",
   "tagline": "StepFun is a Shanghai foundation-model company that has released open-weight language, speech and video models under permissive licenses.",
   "about": [
    "StepFun, formally Shanghai Jieyue Xingchen Intelligent Technology, was founded in April 2023 by Jiang Daxin, Zhu Yibo and Jiao Binxing, all formerly of Microsoft. It is often counted among China's six leading foundation-model startups, and its investors include Tencent and Qiming Venture Partners.",
    "The company builds large language, multimodal, image, video and audio models. Step-2, a trillion-parameter model, appeared in 2024, and it open-sourced Step-Video-T2V and Step-Audio in February 2025 together with Geely.",
    "Its recent open releases, Step 3.5 Flash in February 2026 and Step 3.7 Flash in May 2026, use the Apache 2.0 license. Its latest preview model, Step 5, is closed-weight, so openness varies by release."
   ],
   "mission": "StepFun states an aim of building frontier multimodal models and making many of them available to developers.",
   "notice": null,
   "known_for": [
    "Open Step-Video and Step-Audio models",
    "Step 3.5 Flash mixture-of-experts model",
    "Trillion-parameter Step-2"
   ],
   "focus": [
    "language models",
    "multimodal",
    "speech"
   ],
   "contributions": [
    {
     "title": "Step-2 and multimodal models",
     "note": "StepFun announces the trillion-parameter Step-2, Step-1.5V and Step-1X.",
     "year": "2024"
    },
    {
     "title": "Step-Video-T2V and Step-Audio",
     "note": "Both are open-sourced in February 2025.",
     "year": "2025"
    },
    {
     "title": "Step 3",
     "note": "The company releases Step 3 in July 2025.",
     "year": "2025"
    },
    {
     "title": "Step 3.5 Flash",
     "note": "A 196B-parameter mixture-of-experts model with 11B active parameters is released under Apache 2.0.",
     "year": "2026"
    },
    {
     "title": "Step 3.7 Flash",
     "note": "A further Apache 2.0 release follows in May 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "Step-Video-T2V",
     "url": "https://github.com/stepfun-ai/Step-Video-T2V",
     "note": "Text-to-video model."
    },
    {
     "name": "gelab-zero",
     "url": "https://github.com/stepfun-ai/gelab-zero",
     "note": "GUI agent solution."
    },
    {
     "name": "Step1X-Edit",
     "url": "https://github.com/stepfun-ai/Step1X-Edit",
     "note": "Open image editing model."
    },
    {
     "name": "Step-3.5-Flash",
     "url": "https://github.com/stepfun-ai/Step-3.5-Flash",
     "note": "Fast agentic language model."
    }
   ],
   "hf_highlights": [
    {
     "name": "Step-3.5-Flash",
     "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash",
     "kind": "model",
     "note": "Flagship fast model."
    },
    {
     "name": "GOT-OCR2_0",
     "url": "https://huggingface.co/stepfun-ai/GOT-OCR2_0",
     "kind": "model",
     "note": "OCR model."
    },
    {
     "name": "step3",
     "url": "https://huggingface.co/stepfun-ai/step3",
     "kind": "model",
     "note": "Step-3 model."
    },
    {
     "name": "GEdit-Bench",
     "url": "https://huggingface.co/datasets/stepfun-ai/GEdit-Bench",
     "kind": "dataset",
     "note": "Image editing benchmark."
    }
   ],
   "papers": [
    {
     "title": "Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding",
     "url": "https://arxiv.org/abs/2507.19427",
     "year": "2025"
    },
    {
     "title": "Step-Video-T2V Technical Report: The Practice, Challenges, and Future of Video Foundation Model",
     "url": "https://arxiv.org/abs/2502.10248",
     "year": "2025"
    },
    {
     "title": "General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model",
     "url": "https://arxiv.org/abs/2409.01704",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Step-3.5-Flash",
     "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash",
     "note": "Most liked model on its HF page"
    },
    {
     "name": "Step-5-Preview",
     "url": "https://huggingface.co/stepfun-ai/Step-5-Preview",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights released on Hugging Face and code on GitHub; licence per model.",
   "key_people": [
    {
     "name": "Jiang Daxin",
     "role": "Co-founder",
     "url": "https://www.stepfun.com/",
     "note": null
    },
    {
     "name": "Zhu Yibo",
     "role": "Co-founder",
     "url": "https://www.stepfun.com/",
     "note": null
    },
    {
     "name": "Jiao Binxing",
     "role": "Co-founder",
     "url": "https://www.stepfun.com/",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the official site, GitHub issues or Hugging Face discussions.",
   "sources": [
    "https://huggingface.co/stepfun-ai",
    "https://www.stepfun.com/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "stepfun",
   "mlat": 31.23,
   "mlon": 121.41153,
   "logo": true
  },
  {
   "orig_id": "stockmark",
   "name": "Stockmark",
   "alias": null,
   "full_name": "Stockmark",
   "type": "startup",
   "founded": 2016,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://stockmark.co.jp",
   "huggingface": "https://huggingface.co/stockmark",
   "github": "https://github.com/stockmarkteam",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Tokyo company that releases Japanese document-reading and language models.",
   "tagline": "Stockmark is a Tokyo company that has released large Japanese-language models, including the 100-billion-parameter Stockmark-100b, under permissive licenses.",
   "about": [
    "Stockmark is a Japanese company focused on business information and language models. Its co-founder and CTO is Kosuke Arima, and its researchers publish models on Hugging Face under the stockmark organization.",
    "Stockmark-100b is a 100-billion-parameter model pretrained from scratch on about 910 billion tokens of Japanese and English text and released under the MIT license. The company later released Stockmark-2-100B-Instruct and a multimodal Stockmark-2-VL-100B, both developed under Japan's GENIAC program.",
    "The models target Japanese business use, with a stated focus on reducing hallucination in domain questions."
   ],
   "mission": "Stockmark aims to provide Japanese-first language models that companies can run and adapt, with open weights.",
   "notice": null,
   "known_for": [
    "Stockmark-100b Japanese model",
    "Stockmark-2 100B series",
    "MIT-licensed Japanese LLMs"
   ],
   "focus": [
    "language models",
    "documents",
    "Japanese"
   ],
   "contributions": [
    {
     "title": "Stockmark-100b",
     "note": "A 100B Japanese and English model trained from scratch is published under the MIT license.",
     "year": "2024"
    },
    {
     "title": "Stockmark-2-100B-Instruct",
     "note": "A second-generation 100B model is developed under GENIAC.",
     "year": "2025"
    },
    {
     "title": "Stockmark-2-VL-100B",
     "note": "A vision-language follow-up is released.",
     "year": "2025"
    },
    {
     "title": "Availability on NVIDIA NIM",
     "note": "Stockmark-2-100B-Instruct is offered through NVIDIA's model catalogue.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "bert-book",
     "url": "https://github.com/stockmarkteam/bert-book",
     "note": "Support page for a book on NLP with BERT."
    },
    {
     "name": "ner-wikipedia-dataset",
     "url": "https://github.com/stockmarkteam/ner-wikipedia-dataset",
     "note": "Japanese NER dataset."
    },
    {
     "name": "mind-recommenders-pytorch",
     "url": "https://github.com/stockmarkteam/mind-recommenders-pytorch",
     "note": "Unofficial MIND recommender implementations."
    }
   ],
   "hf_highlights": [
    {
     "name": "gpt-neox-japanese-1.4b",
     "url": "https://huggingface.co/stockmark/gpt-neox-japanese-1.4b",
     "kind": "model",
     "note": "Japanese GPT-NeoX."
    },
    {
     "name": "stockmark-13b",
     "url": "https://huggingface.co/stockmark/stockmark-13b",
     "kind": "model",
     "note": "13B Japanese model."
    },
    {
     "name": "OmniDocBench-JASyn",
     "url": "https://huggingface.co/datasets/stockmark/OmniDocBench-JASyn",
     "kind": "dataset",
     "note": "Japanese doc benchmark."
    },
    {
     "name": "ner-wikipedia-dataset",
     "url": "https://huggingface.co/datasets/stockmark/ner-wikipedia-dataset",
     "kind": "dataset",
     "note": "NER dataset."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Stockmark-DocReasoner-Qwen2.5-VL-32B",
     "url": "https://huggingface.co/stockmark/Stockmark-DocReasoner-Qwen2.5-VL-32B",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [
    {
     "name": "Kosuke Arima",
     "role": "Co-founder and CTO",
     "url": "https://jp.linkedin.com/in/kosuke-arima-428460189",
     "note": null
    },
    {
     "name": "Takahiro Omi",
     "role": "Researcher and model author",
     "url": "https://huggingface.co/stockmark",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website.",
   "sources": [
    "https://huggingface.co/stockmark",
    "https://stockmark.co.jp"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "stockmark",
   "mlat": 35.71062,
   "mlon": 139.62471,
   "logo": true
  },
  {
   "orig_id": "swallow-llm",
   "name": "Swallow LLM",
   "alias": null,
   "full_name": "Swallow LLM",
   "type": "academic",
   "founded": 2023,
   "country": "Japan",
   "city": "Tokyo",
   "lat": 35.68,
   "lon": 139.69,
   "website": "https://swallow-llm.github.io/",
   "huggingface": "https://huggingface.co/tokyotech-llm",
   "github": "https://github.com/swallow-llm",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "LLM project led by the Okazaki and Yokota labs at the Institute of Science Tokyo, adapting open models for Japanese.",
   "tagline": "Swallow is a Japanese LLM project at the Institute of Science Tokyo that continually pretrains open models to give them strong Japanese ability.",
   "about": [
    "Swallow is led by researchers in the Okazaki and Yokota laboratories at the Institute of Science Tokyo, together with the National Institute of Advanced Industrial Science and Technology (AIST). Naoaki Okazaki leads the project and the pre-training corpus team, Rio Yokota leads LLM training, and Sakae Mizuki leads instruction tuning and evaluation.",
    "The project adapts existing open models such as Llama, Mistral, Gemma, Qwen and gpt-oss to Japanese through continual pretraining and instruction tuning. Its model list runs from the first Swallow models to Llama 3.3 Swallow 70B, Qwen3 Swallow and GPT-OSS Swallow.",
    "Weights are published on Hugging Face under the tokyotech-llm organization, and the project documents its corpus and evaluation work for the Japanese community."
   ],
   "mission": "Swallow aims to build and openly release Japanese large language models and the corpora and evaluation methods needed to improve them.",
   "notice": null,
   "known_for": [
    "Japanese continual-pretrained Llama models",
    "Swallow corpus and evaluation work",
    "Llama 3.3 Swallow 70B"
   ],
   "focus": [
    "language models",
    "Japanese",
    "medical"
   ],
   "contributions": [
    {
     "title": "First Swallow models",
     "note": "Llama-based Swallow models extend Japanese ability of open LLMs.",
     "year": "2023"
    },
    {
     "title": "Llama 3 Swallow",
     "note": "Models built on Llama 3 are released.",
     "year": "2024"
    },
    {
     "title": "Llama 3.1 Swallow",
     "note": "The line is updated to Llama 3.1.",
     "year": "2024"
    },
    {
     "title": "Llama 3.3 Swallow 70B",
     "note": "A 70B model is added to the family.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "swallow-evaluation-instruct",
     "url": "https://github.com/swallow-llm/swallow-evaluation-instruct",
     "note": "Evaluation framework for post-trained models."
    },
    {
     "name": "swallow-evaluation",
     "url": "https://github.com/swallow-llm/swallow-evaluation",
     "note": "Evaluation scripts."
    },
    {
     "name": "swallow-corpus",
     "url": "https://github.com/swallow-llm/swallow-corpus",
     "note": "Corpus code."
    },
    {
     "name": "doubri",
     "url": "https://github.com/swallow-llm/doubri",
     "note": "Deduplication tool."
    }
   ],
   "hf_highlights": [
    {
     "name": "Llama-3-Swallow-8B-Instruct-v0.1",
     "url": "https://huggingface.co/tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1",
     "kind": "model",
     "note": "Japanese Llama 3."
    },
    {
     "name": "Llama-3.1-Swallow-8B-Instruct-v0.5",
     "url": "https://huggingface.co/tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.5",
     "kind": "model",
     "note": "Japanese Llama 3.1."
    },
    {
     "name": "swallow-code-v2",
     "url": "https://huggingface.co/datasets/tokyotech-llm/swallow-code-v2",
     "kind": "dataset",
     "note": "Code dataset."
    },
    {
     "name": "swallow-math-v2",
     "url": "https://huggingface.co/datasets/tokyotech-llm/swallow-math-v2",
     "kind": "dataset",
     "note": "Math dataset."
    }
   ],
   "papers": [
    {
     "title": "Continual Pre-Training for Cross-Lingual LLM Adaptation: Enhancing Japanese Language Capabilities",
     "url": "https://arxiv.org/abs/2404.17790",
     "year": "2024"
    },
    {
     "title": "Building a Large Japanese Web Corpus for Large Language Models",
     "url": "https://arxiv.org/abs/2404.17733",
     "year": "2024"
    },
    {
     "title": "Rewriting Pre-Training Data Boosts LLM Performance in Math and Code",
     "url": "https://arxiv.org/abs/2505.02881",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "GPT-OSS-Swallow-120B-RL-v0.1",
     "url": "https://huggingface.co/tokyotech-llm/GPT-OSS-Swallow-120B-RL-v0.1",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Medical-Qwen3-Swallow-8B",
     "url": "https://huggingface.co/tokyotech-llm/Medical-Qwen3-Swallow-8B",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face, code on GitHub.",
   "key_people": [
    {
     "name": "Naoaki Okazaki",
     "role": "Professor; project leader",
     "url": "https://swallow-llm.github.io/index.en.html",
     "note": null
    },
    {
     "name": "Rio Yokota",
     "role": "Professor; LLM training team leader",
     "url": "https://swallow-llm.github.io/index.en.html",
     "note": null
    },
    {
     "name": "Sakae Mizuki",
     "role": "Researcher; instruction tuning and evaluation leader",
     "url": "https://swallow-llm.github.io/index.en.html",
     "note": null
    },
    {
     "name": "Hiroya Takamura",
     "role": "AIST team leader; project manager",
     "url": "https://swallow-llm.github.io/index.en.html",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the project website.",
   "sources": [
    "https://huggingface.co/tokyotech-llm",
    "https://swallow-llm.github.io/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "swallow-llm",
   "mlat": 35.64938,
   "mlon": 139.62471,
   "logo": true
  },
  {
   "orig_id": "swap-uniba",
   "name": "SWAP Research Group",
   "alias": "University of Bari",
   "full_name": "SWAP Research Group (University of Bari)",
   "type": "academic",
   "founded": null,
   "country": "Italy",
   "city": "Bari",
   "lat": 41.1171,
   "lon": 16.8719,
   "website": "https://github.com/swapUniba",
   "huggingface": "https://huggingface.co/swap-uniba",
   "github": "https://github.com/swapUniba",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Semantic Web Access and Personalization group at the University of Bari, creators of the LLaMAntino Italian language models.",
   "tagline": "Bari's SWAP group, creators of the LLaMAntino family of open Italian language models.",
   "about": [
    "The Semantic Web Access and Personalization group works at the University of Bari. It is known for LLaMAntino, an Italian adaptation of Meta's Llama models.",
    "The first paper, in December 2023, adapted LLaMA 2 to Italian. The ANITA models followed in 2024 on Llama 3, trained with the Leonardo supercomputer at CINECA.",
    "In 2025 a model was reported developed with Treccani."
   ],
   "mission": "The group aims to provide effective open language models for Italian.",
   "notice": null,
   "known_for": [
    "LLaMAntino Italian models",
    "ANITA Llama-3 Italian",
    "Open weights for commercial use"
   ],
   "focus": [
    "Italian LLMs",
    "recommender systems"
   ],
   "contributions": [
    {
     "title": "LLaMAntino paper",
     "note": "LLaMA 2 models for Italian released.",
     "year": "2023"
    },
    {
     "title": "LLaMAntino-3-ANITA",
     "note": "Llama 3 based model paper in May.",
     "year": "2024"
    },
    {
     "title": "Treccani collaboration",
     "note": "Model with Treccani and Puglia Region reported.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "LLaMAntino",
     "url": "https://github.com/swapUniba/LLaMAntino",
     "note": "Open-source repository. (39 stars)"
    },
    {
     "name": "ClayRS",
     "url": "https://github.com/swapUniba/ClayRS",
     "note": "Complexly represent contents, build recommender systems, evaluate them. All in one place! (37 stars)"
    },
    {
     "name": "ConvRecSysDataset",
     "url": "https://github.com/swapUniba/ConvRecSysDataset",
     "note": "Open-source repository. (9 stars)"
    },
    {
     "name": "multimodal_ml1m_dbbook_lfm2k",
     "url": "https://github.com/swapUniba/multimodal_ml1m_dbbook_lfm2k",
     "note": "Open-source repository. (8 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA",
     "url": "https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA",
     "kind": "model",
     "note": "Model with 31 likes on the Hub."
    },
    {
     "name": "swap-uniba/LLaMAntino-2-chat-13b-hf-UltraChat-ITA",
     "url": "https://huggingface.co/swap-uniba/LLaMAntino-2-chat-13b-hf-UltraChat-ITA",
     "kind": "model",
     "note": "Model with 16 likes on the Hub."
    },
    {
     "name": "swap-uniba/itwiki-march-2024",
     "url": "https://huggingface.co/datasets/swap-uniba/itwiki-march-2024",
     "kind": "dataset",
     "note": "Dataset with 2 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Advanced Natural-based interaction for the ITAlian language: LLaMAntino-3-ANITA",
     "url": "https://arxiv.org/abs/2405.07101",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "LLaMAntino-3-ANITA",
     "url": "https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA",
     "note": "Italian 8B instruct model"
    },
    {
     "name": "LLaMAntino-2 chat 13B",
     "url": "https://huggingface.co/swap-uniba/LLaMAntino-2-chat-13b-hf-UltraChat-ITA",
     "note": "Italian chat model"
    },
    {
     "name": "ClayRS",
     "url": "https://github.com/swapUniba/ClayRS",
     "note": "Recommender systems library"
    },
    {
     "name": "LLaMAntino",
     "url": "https://github.com/swapUniba/LLaMAntino",
     "note": "Project repository"
    }
   ],
   "openness": "Models are public on Hugging Face and code on GitHub; model cards follow Llama licences.",
   "key_people": [
    {
     "name": "Marco Polignano",
     "role": "Developer of LLaMAntino",
     "url": "https://github.com/marcopoli",
     "note": null
    },
    {
     "name": "Pierpaolo Basile",
     "role": "Researcher",
     "url": "https://arxiv.org/abs/2312.09993",
     "note": null
    },
    {
     "name": "Giovanni Semeraro",
     "role": "Professor",
     "url": "https://arxiv.org/abs/2312.09993",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://github.com/swapUniba",
    "https://huggingface.co/swap-uniba",
    "https://github.com/swapUniba",
    "https://arxiv.org/abs/2405.07101"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "swap-research-group",
   "logo": true
  },
  {
   "orig_id": "swiss-ai-initiative",
   "name": "Swiss AI Initiative",
   "alias": null,
   "full_name": "Swiss AI Initiative",
   "type": "institute",
   "founded": 2023,
   "country": "Switzerland",
   "city": "Zurich",
   "lat": 47.3769,
   "lon": 8.5417,
   "website": "https://www.swiss-ai.org",
   "huggingface": "https://huggingface.co/swiss-ai",
   "github": "https://github.com/swiss-ai",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "National initiative of ETH Zurich and EPFL, with partners, that trained the Apertus open language models on the Alps supercomputer. It releases weights, data recipes and training code.",
   "tagline": "The Swiss AI Initiative is a national ETH Zurich and EPFL effort that trained the fully open Apertus language models on the Alps supercomputer.",
   "about": [
    "The Swiss AI Initiative started in December 2023, seeded with more than 10 million GPU hours on the Alps system run by CSCS and a CHF 20 million grant from the ETH Domain. It is run with the ETH AI Center and the EPFL AI Center, with Swisscom as strategic partner.",
    "Its main output is Apertus, released on 2 September 2025 in 8B and 70B sizes under Apache 2.0. Apertus was trained on over 1,800 languages and designed to follow European copyright law, with an emphasis on transparency and auditability.",
    "An Apertus 1.5 release is listed for July 2026."
   ],
   "mission": "The initiative aims to build transparent, sovereign foundation models for Switzerland and to share them with the wider research community.",
   "notice": null,
   "known_for": [
    "Apertus open language models",
    "Alps supercomputer training",
    "Transparency and copyright-aware data"
   ],
   "focus": [
    "language models",
    "multilingual",
    "datasets",
    "training infrastructure"
   ],
   "contributions": [
    {
     "title": "Initiative launched",
     "note": "The Swiss AI Initiative begins with over 10 million GPU hours and a CHF 20 million grant.",
     "year": "2023"
    },
    {
     "title": "Apertus 1",
     "note": "8B and 70B models released under Apache 2.0 on 2 September.",
     "year": "2025"
    },
    {
     "title": "Apertus 1.5",
     "note": "A follow-up release is listed for July 2026.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "pretrain-data",
     "url": "https://github.com/swiss-ai/pretrain-data",
     "note": "Pretraining data reconstruction scripts for Apertus (170 stars)"
    },
    {
     "name": "apertus-tech-report",
     "url": "https://github.com/swiss-ai/apertus-tech-report",
     "note": "Tech Report of the Apertus LLM (144 stars)"
    },
    {
     "name": "Megatron-LM",
     "url": "https://github.com/swiss-ai/Megatron-LM",
     "note": "Ongoing research training transformer models at scale (53 stars)"
    },
    {
     "name": "ESFM",
     "url": "https://github.com/swiss-ai/ESFM",
     "note": "Open-source repository (47 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "swiss-ai/Apertus-8B-Instruct-2509",
     "url": "https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509",
     "kind": "model",
     "note": "Model: 191,142 downloads, 505 likes."
    },
    {
     "name": "swiss-ai/Apertus-8B-2509",
     "url": "https://huggingface.co/swiss-ai/Apertus-8B-2509",
     "kind": "model",
     "note": "Model: 10,421 downloads, 175 likes."
    },
    {
     "name": "swiss-ai/nasa",
     "url": "https://huggingface.co/datasets/swiss-ai/nasa",
     "kind": "dataset",
     "note": "Dataset: 3,321 downloads, 2 likes."
    },
    {
     "name": "swiss-ai/apertus-mini-webgpu",
     "url": "https://huggingface.co/spaces/swiss-ai/apertus-mini-webgpu",
     "kind": "space",
     "note": "Space: 12 likes on the Hub."
    }
   ],
   "papers": [
    {
     "title": "Apertus: Democratizing Open and Compliant LLMs for Global Language Environments",
     "url": "https://arxiv.org/abs/2509.14233",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Apertus-8B-Instruct",
     "url": "https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509",
     "note": "Fully open 8B instruction model"
    },
    {
     "name": "Apertus-70B-Instruct",
     "url": "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509",
     "note": "Fully open 70B instruction model"
    },
    {
     "name": "pretrain-data",
     "url": "https://github.com/swiss-ai/pretrain-data",
     "note": "Scripts to reconstruct the Apertus pretraining data"
    },
    {
     "name": "Apertus tech report",
     "url": "https://github.com/swiss-ai/apertus-tech-report",
     "note": "Technical report source"
    }
   ],
   "openness": "Weights, training recipes, data-reconstruction scripts and the technical report are public; code is Apache-2.0.",
   "key_people": [
    {
     "name": "Antoine Bosselut",
     "role": "Co-lead of the Apertus project, EPFL",
     "url": "https://atcbosselut.github.io/",
     "note": "Senior author of the Apertus report."
    },
    {
     "name": "Martin Jaggi",
     "role": "Professor, EPFL; Apertus senior author",
     "url": "https://arxiv.org/abs/2509.14233",
     "note": "Listed among senior authors of the Apertus report."
    },
    {
     "name": "Imanol Schlag",
     "role": "Researcher, ETH AI Center; Apertus senior author",
     "url": "https://ischlag.github.io/",
     "note": "Listed among senior authors of the Apertus report."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the website.",
   "sources": [
    "https://www.swiss-ai.org",
    "https://huggingface.co/swiss-ai",
    "https://github.com/swiss-ai",
    "https://arxiv.org/abs/2509.14233"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "swiss-ai-initiative",
   "mlat": 47.39855,
   "mlon": 8.48632,
   "logo": true
  },
  {
   "orig_id": "taide",
   "name": "TAIDE",
   "alias": null,
   "full_name": "TAIDE",
   "type": "institute",
   "founded": 2023,
   "country": "Taiwan",
   "city": "Taipei",
   "lat": 25.03,
   "lon": 121.57,
   "website": "https://taide.tw/",
   "huggingface": "https://huggingface.co/taide",
   "github": "https://github.com/taide-taiwan",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Taiwan trustworthy AI dialogue engine project releasing Traditional Chinese models. Government-initiated.",
   "tagline": "TAIDE is a Taiwanese government-backed project releasing open Traditional Chinese language models for trustworthy local use.",
   "about": [
    "TAIDE, the Trustworthy AI Dialogue Engine, launched on 27 April 2023. The National Science and Technology Council acts as guiding agency, and the National Institutes of Applied Research implements the work.",
    "The project has released the TAIDE-LX-7B model in 2024, Llama 3 and Llama 3.1 based 8B chat models, and Gemma 3 based 12B chat models in 2025 and 2026. A small embedding model, Embeddinggemma-GTAIDE-300m, appeared in 2026.",
    "Its models are intended to serve Traditional Chinese users and Taiwanese institutions."
   ],
   "mission": "TAIDE aims to provide a trustworthy Traditional Chinese language model base for Taiwan's public and private sectors.",
   "notice": null,
   "known_for": [
    "Traditional Chinese LLMs",
    "Government-sponsored open models",
    "Llama and Gemma derived chat models"
   ],
   "focus": [
    "language models",
    "Traditional Chinese"
   ],
   "contributions": [
    {
     "title": "Project launch",
     "note": "The National Science and Technology Council starts TAIDE on 27 April.",
     "year": "2023"
    },
    {
     "title": "TAIDE-LX-7B",
     "note": "First model of the family published; a Llama 3 based 8B preview followed.",
     "year": "2024"
    },
    {
     "title": "Llama3.1-TAIDE-LX-8B-Chat",
     "note": "8B chat model with a 131K context length.",
     "year": "2025"
    },
    {
     "title": "Gemma-3-TAIDE-12b-Chat",
     "note": "12B chat model based on Gemma 3.",
     "year": "2025"
    },
    {
     "title": "Embeddinggemma-GTAIDE-300m",
     "note": "Embedding model aimed at retrieval of Taiwanese legal regulations.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "taide-bench-eval",
     "url": "https://github.com/taide-taiwan/taide-bench-eval",
     "note": "Benchmark evaluation code."
    },
    {
     "name": "TaiwanVQA",
     "url": "https://github.com/taide-taiwan/TaiwanVQA",
     "note": "Taiwan visual QA."
    }
   ],
   "hf_highlights": [
    {
     "name": "taide/Llama-3.1-TAIDE-LX-8B-Chat",
     "url": "https://huggingface.co/taide/Llama-3.1-TAIDE-LX-8B-Chat",
     "kind": "model",
     "note": "Most downloaded."
    },
    {
     "name": "taide/Gemma-3-TAIDE-12b-Chat",
     "url": "https://huggingface.co/taide/Gemma-3-TAIDE-12b-Chat",
     "kind": "model",
     "note": "12B chat."
    },
    {
     "name": "taide/embeddinggemma-GTAIDE-300m-2605",
     "url": "https://huggingface.co/taide/embeddinggemma-GTAIDE-300m-2605",
     "kind": "model",
     "note": "Embeddings."
    },
    {
     "name": "taide/taide-bench",
     "url": "https://huggingface.co/datasets/taide/taide-bench",
     "kind": "dataset",
     "note": "Benchmark."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Gemma-3-TAIDE-12b-Chat",
     "url": "https://huggingface.co/taide/Gemma-3-TAIDE-12b-Chat",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Llama-3.1-TAIDE-LX-8B-Chat",
     "url": "https://huggingface.co/taide/Llama-3.1-TAIDE-LX-8B-Chat",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [],
   "discord_note": null,
   "contact_note": "taide.tw has a team and news pages with contact options.",
   "sources": [
    "https://huggingface.co/taide",
    "https://taide.tw/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "taide",
   "mlat": 25.06536,
   "mlon": 121.57,
   "logo": true
  },
  {
   "orig_id": "takelab",
   "name": "TakeLab",
   "alias": "University of Zagreb",
   "full_name": "TakeLab (University of Zagreb)",
   "type": "academic",
   "founded": null,
   "country": "Croatia",
   "city": "Zagreb",
   "lat": 45.815,
   "lon": 15.9819,
   "website": "http://takelab.fer.hr",
   "huggingface": null,
   "github": "https://github.com/takelab",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Text analysis and knowledge engineering lab at the Faculty of Electrical Engineering and Computing, University of Zagreb.",
   "tagline": "Zagreb's text analysis and knowledge engineering lab, a leading Croatian NLP group.",
   "about": [
    "TakeLab is a research group at the Faculty of Electrical Engineering and Computing, University of Zagreb, focusing on AI, machine learning and NLP.",
    "Professor Jan Snajder is a prominent member. Work includes Croatian morphological analysis and SemEval shared task entries.",
    "Code is published on GitHub."
   ],
   "mission": "The lab advances AI and NLP research, with attention to Croatian.",
   "notice": null,
   "known_for": [
    "Croatian NLP research",
    "SemEval entries",
    "Text analysis"
   ],
   "focus": [
    "NLP",
    "Croatian",
    "data loading"
   ],
   "contributions": [
    {
     "title": "Inflectional paradigm guessing for Croatian",
     "note": "Talk on morphology.",
     "year": "2012"
    },
    {
     "title": "Hyperpartisan news detection at SemEval",
     "note": "Shared task paper.",
     "year": "2019"
    }
   ],
   "github_highlights": [
    {
     "name": "spacy-udpipe",
     "url": "https://github.com/TakeLab/spacy-udpipe",
     "note": "spaCy + UDPipe (167 stars)"
    },
    {
     "name": "podium",
     "url": "https://github.com/TakeLab/podium",
     "note": "Podium: a framework agnostic Python NLP library for data loading and preprocessing (60 stars)"
    },
    {
     "name": "repetition-sl",
     "url": "https://github.com/TakeLab/repetition-sl",
     "note": "Open-source repository. (5 stars)"
    },
    {
     "name": "Coral",
     "url": "https://github.com/TakeLab/Coral",
     "note": "Coral - Corpus aligner (2 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "spacy-udpipe",
     "url": "https://github.com/takelab/spacy-udpipe",
     "note": "spaCy and UDPipe integration"
    },
    {
     "name": "Podium",
     "url": "https://github.com/takelab/podium",
     "note": "Framework-agnostic NLP data loading library"
    },
    {
     "name": "Coral",
     "url": "https://github.com/takelab/Coral",
     "note": "Corpus aligner"
    }
   ],
   "openness": "Several libraries and tools are open on GitHub.",
   "key_people": [
    {
     "name": "Jan Snajder",
     "role": "Professor, head of the group",
     "url": "https://www.fer.unizg.hr/en/jan.snajder",
     "note": "Leads Croatian NLP work."
    },
    {
     "name": "Martin Tutek",
     "role": "Researcher",
     "url": "https://www.fer.unizg.hr/en/martin.tutek",
     "note": "NLP researcher at the group."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the faculty website.",
   "sources": [
    "http://takelab.fer.hr",
    "https://github.com/takelab"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "takelab",
   "logo": true
  },
  {
   "orig_id": "taltech-nlp",
   "name": "TalTech Language Technology",
   "alias": "TalTechNLP",
   "full_name": "TalTech Language Technology (TalTechNLP)",
   "type": "academic",
   "founded": null,
   "country": "Estonia",
   "city": "Tallinn",
   "lat": 59.437,
   "lon": 24.7536,
   "website": "https://taltech.ee",
   "huggingface": "https://huggingface.co/TalTechNLP",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Tallinn University of Technology language technology group publishing Estonian speech and language models on Hugging Face.",
   "tagline": "The TalTech language technology lab publishing open Estonian speech recognition and translation models.",
   "about": [
    "The Laboratory of Language Technology at Tallinn University of Technology develops speech and language technology for Estonian. Tanel Alumae leads related research.",
    "Its general-purpose Estonian ASR is a Whisper-large-v2 fine-tune on about 1200 hours, reporting about 12 percent word error rate on Common Voice 11.",
    "A Whisper fine-tune for Estonian-English-Russian speech translation was published in 2024, and children's speech recognition work in 2025."
   ],
   "mission": "The lab builds practical Estonian speech technology and shares models openly.",
   "notice": null,
   "known_for": [
    "Estonian Whisper ASR",
    "Estonian speech translation",
    "TalTech Estonian Speech Dataset"
   ],
   "focus": [
    "Estonian",
    "speech recognition",
    "language identification"
   ],
   "contributions": [
    {
     "title": "Estonian Whisper large model update",
     "note": "Long-segment version dated 3 October.",
     "year": "2023"
    },
    {
     "title": "Estonian conversational speech translation",
     "note": "arXiv paper on fine-tuned end-to-end models.",
     "year": "2024"
    },
    {
     "title": "Children's speech recognition",
     "note": "SLaTE paper by Lilles, Mannik and Alumae.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "TalTechNLP/voxlingua107-epaca-tdnn",
     "url": "https://huggingface.co/TalTechNLP/voxlingua107-epaca-tdnn",
     "kind": "model",
     "note": "Model with 33 likes on the Hub."
    },
    {
     "name": "TalTechNLP/whisper-large-et",
     "url": "https://huggingface.co/TalTechNLP/whisper-large-et",
     "kind": "model",
     "note": "Model with 8 likes on the Hub."
    },
    {
     "name": "TalTechNLP/voxlingua107_wds",
     "url": "https://huggingface.co/datasets/TalTechNLP/voxlingua107_wds",
     "kind": "dataset",
     "note": "Dataset with 5 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "VoxLingua107 ECAPA",
     "url": "https://huggingface.co/TalTechNLP/voxlingua107-epaca-tdnn",
     "note": "Spoken language identification model"
    },
    {
     "name": "Whisper Estonian",
     "url": "https://huggingface.co/TalTechNLP/whisper-large-et",
     "note": "Estonian Whisper model"
    },
    {
     "name": "voxlingua107_wds",
     "url": "https://huggingface.co/datasets/TalTechNLP/voxlingua107_wds",
     "note": "Language identification dataset"
    }
   ],
   "openness": "Models and datasets are public on the TalTechNLP Hugging Face organisation.",
   "key_people": [
    {
     "name": "Tanel Alumäe",
     "role": "Researcher, speech technology",
     "url": "https://arxiv.org/abs/2407.03809",
     "note": "Leads Estonian ASR and speech translation models."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the university website.",
   "sources": [
    "https://taltech.ee",
    "https://huggingface.co/TalTechNLP"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "taltech-language-technology",
   "logo": true
  },
  {
   "orig_id": "tartunlp",
   "name": "TartuNLP",
   "alias": "University of Tartu",
   "full_name": "TartuNLP (University of Tartu)",
   "type": "academic",
   "founded": null,
   "country": "Estonia",
   "city": "Tartu",
   "lat": 58.378,
   "lon": 26.729,
   "website": "https://tartunlp.ai",
   "huggingface": "https://huggingface.co/tartuNLP",
   "github": "https://github.com/TartuNLP",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "NLP group at the University of Tartu that publishes Estonian and Finno-Ugric language models, speech synthesis and translation tools.",
   "tagline": "TartuNLP is the University of Tartu's language technology group, building Estonian and Finno-Ugric models, translation and speech tools.",
   "about": [
    "TartuNLP is a natural language processing group at the University of Tartu in Estonia. Its site presents work on Estonian language models, machine translation and speech synthesis.",
    "Named projects include Llammas, an Estonian-adapted Llama model, EstBERT, and the Neurotõlge translation service. Code and models are published through the TartuNLP GitHub organization."
   ],
   "mission": "The group works to provide open language technology for Estonian and related smaller languages.",
   "notice": null,
   "known_for": [
    "Estonian language models",
    "Neurotõlge machine translation",
    "Finno-Ugric language tools"
   ],
   "focus": [
    "Estonian",
    "speech",
    "machine translation",
    "language models"
   ],
   "contributions": [
    {
     "title": "EstBERT",
     "note": "An Estonian BERT model is published.",
     "year": "2020"
    },
    {
     "title": "Llammas",
     "note": "An Estonian instruction-tuned Llama model is released.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "text-to-speech-worker",
     "url": "https://github.com/TartuNLP/text-to-speech-worker",
     "note": "Estonian multi-speaker neural text-to-speech worker (20 stars)"
    },
    {
     "name": "m2m-100-finetune",
     "url": "https://github.com/TartuNLP/m2m-100-finetune",
     "note": "Scripts for finetuning M2M-100 models (19 stars)"
    },
    {
     "name": "text-to-speech-api",
     "url": "https://github.com/TartuNLP/text-to-speech-api",
     "note": "API for neural text-to-speech synthesis (17 stars)"
    },
    {
     "name": "translation-api",
     "url": "https://github.com/TartuNLP/translation-api",
     "note": "REST API for neural machine translation (14 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "tartuNLP/EstBERT",
     "url": "https://huggingface.co/tartuNLP/EstBERT",
     "kind": "model",
     "note": "Estonian BERT model."
    },
    {
     "name": "tartuNLP/Llama-3.1-EstLLM-70B-Instruct-0826",
     "url": "https://huggingface.co/tartuNLP/Llama-3.1-EstLLM-70B-Instruct-0826",
     "kind": "model",
     "note": "70B Estonian instruction model."
    },
    {
     "name": "tartuNLP/finno-ugric-benchmark",
     "url": "https://huggingface.co/datasets/tartuNLP/finno-ugric-benchmark",
     "kind": "dataset",
     "note": "Benchmark for Finno-Ugric languages."
    },
    {
     "name": "tartuNLP/XTTSv2-est",
     "url": "https://huggingface.co/spaces/tartuNLP/XTTSv2-est",
     "kind": "space",
     "note": "Estonian XTTS speech synthesis demo."
    }
   ],
   "papers": [
    {
     "title": "EstBERT: A Pretrained Language-Specific BERT for Estonian",
     "url": "https://arxiv.org/abs/2011.04784",
     "year": "2020"
    },
    {
     "title": "EstLLM: Enhancing Estonian Capabilities in Multilingual LLMs via Continued Pretraining and Post-Training",
     "url": "https://arxiv.org/abs/2603.02041",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "EstLLM Llama-3.1-8B Instruct",
     "url": "https://huggingface.co/tartuNLP/Llama-3.1-EstLLM-8B-Instruct-1125",
     "note": "Estonian-adapted LLM"
    },
    {
     "name": "Llammas",
     "url": "https://huggingface.co/tartuNLP/Llammas",
     "note": "Estonian instruction model"
    },
    {
     "name": "text-to-speech-api",
     "url": "https://github.com/TartuNLP/text-to-speech-api",
     "note": "Estonian TTS API"
    },
    {
     "name": "EstBERT",
     "url": "https://huggingface.co/tartuNLP/EstBERT",
     "note": "Estonian BERT"
    }
   ],
   "openness": "Models and datasets are public on Hugging Face; code on GitHub is mostly MIT.",
   "key_people": [
    {
     "name": "Mark Fišel",
     "role": "Associate professor, head of the group",
     "url": "https://arxiv.org/abs/2603.02041",
     "note": "Author of the EstLLM report."
    },
    {
     "name": "Kairit Sirts",
     "role": "Associate professor",
     "url": "https://arxiv.org/abs/2603.02041",
     "note": "Co-author of EstBERT and EstLLM."
    },
    {
     "name": "Tanel Alumäe",
     "role": "Researcher (also TalTech)",
     "url": "https://arxiv.org/abs/2603.02041",
     "note": "Speech technology co-author on EstLLM."
    },
    {
     "name": "Taido Purason",
     "role": "Researcher",
     "url": "https://arxiv.org/abs/2603.02041",
     "note": "Co-author of EstLLM and Llammas-related work."
    }
   ],
   "discord_note": null,
   "contact_note": "The group's website and GitHub organisation are the public entry points; its site could not be read as plain text.",
   "sources": [
    "https://tartunlp.ai",
    "https://huggingface.co/tartuNLP",
    "https://github.com/TartuNLP"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "tartunlp",
   "logo": true
  },
  {
   "orig_id": "te-hiku-media",
   "name": "Te Hiku Media",
   "alias": null,
   "full_name": "Te Hiku Media",
   "type": "nonprofit",
   "founded": null,
   "country": "New Zealand",
   "city": "Kaitaia",
   "lat": -35.1133,
   "lon": 173.263,
   "website": "https://tehiku.nz",
   "huggingface": null,
   "github": "https://github.com/TeHikuMedia",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@tehiku.co.nz",
   "summary": "Te Hiku Media is a Māori iwi media organisation in Kaitaia that builds te reo Māori speech technology under Māori data sovereignty. Its code and a data-sovereignty licence are public on GitHub; some data and models are governed by that licence rather than open licences.",
   "tagline": "Te Hiku Media is a Māori iwi media organization in Kaitaia that builds te reo Māori speech technology under Māori data sovereignty.",
   "about": [
    "Te Hiku Media, formally Te Reo Irirangi o Te Hiku o Te Ika, is an iwi radio and media organization in Kaitaia whose site dates its roots to 1991. It runs Te Hiku Tech, which builds tools for te reo Māori.",
    "Its Kōrero Māori campaign collected labelled speech under the Kaitiakitanga License, which restricts use of the data to the benefit of Māori people. The organization reported a speech recognizer for te reo Māori built with DeepSpeech, with accuracy of about 92 percent.",
    "All content on its site is covered by the same licence, and contributors retain ownership of their data."
   ],
   "mission": "Te Hiku Media aims to revitalize te reo Māori with technology that Māori communities own and control.",
   "notice": null,
   "known_for": [
    "te reo Māori speech recognition",
    "Kaitiakitanga License",
    "Indigenous data sovereignty"
   ],
   "focus": [
    "speech",
    "data sovereignty",
    "low-resource languages"
   ],
   "contributions": [
    {
     "title": "Kōrero Māori campaign",
     "note": "Crowdsourced speech collection for te reo Māori under the Kaitiakitanga License.",
     "year": "2018"
    },
    {
     "title": "NeurIPS 2020 keynote",
     "note": "Leaders present indigenous speech recognition and data sovereignty.",
     "year": "2020"
    },
    {
     "title": "Whare Kōrero platform",
     "note": "A platform hosting more than 30 years of digitised archive material.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "Kaitiakitanga-License",
     "url": "https://github.com/TeHikuMedia/Kaitiakitanga-License",
     "note": "License."
    },
    {
     "name": "nga-tautohetohe-reo",
     "url": "https://github.com/TeHikuMedia/nga-tautohetohe-reo",
     "note": "Corpus."
    },
    {
     "name": "korero-maori-demos",
     "url": "https://github.com/TeHikuMedia/korero-maori-demos",
     "note": "Demos."
    },
    {
     "name": "wahi-korero",
     "url": "https://github.com/TeHikuMedia/wahi-korero",
     "note": "Audio segment tools."
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "Kaitiakitanga License",
     "url": "https://github.com/TeHikuMedia/Kaitiakitanga-License",
     "note": "Data sovereignty licence"
    },
    {
     "name": "nga-tautohetohe-reo",
     "url": "https://github.com/TeHikuMedia/nga-tautohetohe-reo",
     "note": "Te reo corpus from Hansard"
    },
    {
     "name": "korero-maori-demos",
     "url": "https://github.com/TeHikuMedia/korero-maori-demos",
     "note": "Kōrero Māori API demos"
    }
   ],
   "openness": "Code and corpora on GitHub; data held under Māori data sovereignty principles, so openness is conditional. Hugging Face presence not found. Borderline inclusion.",
   "key_people": [
    {
     "name": "Peter-Lucas Jones",
     "role": "Chief Executive",
     "url": "https://neurips.cc/virtual/2020/20652",
     "note": "Also chair of the National Māori Radio Network."
    },
    {
     "name": "Keoni Mahelona",
     "role": "Chief Technology Officer",
     "url": "https://neurips.cc/virtual/2020/20652",
     "note": "Leads Te Hiku Tech."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://tehiku.nz); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://tehiku.nz",
    "https://github.com/TeHikuMedia"
   ],
   "confidence": "high",
   "region": "Oceania",
   "id": "te-hiku-media",
   "logo": true
  },
  {
   "orig_id": "tii-falcon",
   "name": "Technology Innovation Institute",
   "alias": "Falcon",
   "full_name": "Technology Innovation Institute (Falcon)",
   "type": "institute",
   "founded": 2020,
   "country": "United Arab Emirates",
   "city": "Abu Dhabi",
   "lat": 24.45,
   "lon": 54.38,
   "website": "https://www.tii.ae/",
   "huggingface": "https://huggingface.co/tiiuae",
   "github": "https://github.com/tiiuae",
   "x": "https://twitter.com/TIIuae",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Research centre of the Abu Dhabi government's Advanced Technology Research Council, publisher of the Falcon models. Government-funded.",
   "tagline": "TII is Abu Dhabi's government-funded research institute and the publisher of the Falcon open language models.",
   "about": [
    "The Technology Innovation Institute is part of the Advanced Technology Research Council of the Abu Dhabi government. It works across AI, quantum computing, robotics, cryptography and other fields, and its Artificial Intelligence and Digital Science Research Center builds the Falcon models.",
    "Falcon 40B and 7B were released in 2023, followed by Falcon 180B in September 2023, Falcon 2 11B in May 2024 and Falcon 3 in December 2024. Later work includes the hybrid Falcon-H1 family and Falcon-H1 Arabic in January 2026.",
    "TII is state funded, so its openness follows its license choices per release."
   ],
   "mission": "TII aims to build sovereign, efficient AI models and release them for the global research community.",
   "notice": null,
   "known_for": [
    "Falcon open LLM family",
    "Falcon 180B",
    "Falcon-H1 hybrid models"
   ],
   "focus": [
    "language models",
    "vision",
    "1-bit models"
   ],
   "contributions": [
    {
     "title": "Falcon 40B and 7B",
     "note": "Open weights released.",
     "year": "2023"
    },
    {
     "title": "Falcon 180B",
     "note": "Released in September.",
     "year": "2023"
    },
    {
     "title": "Falcon 2 11B",
     "note": "Second generation with a vision-language variant, in May.",
     "year": "2024"
    },
    {
     "title": "Falcon 3",
     "note": "Five base models and instruction variants released 17 December.",
     "year": "2024"
    },
    {
     "title": "Falcon-H1 Arabic",
     "note": "Released in January with leading Arabic leaderboard results.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "jamrf",
     "url": "https://github.com/tiiuae/jamrf",
     "note": "SDR-based jamming with GNU Radio."
    },
    {
     "name": "rclgo",
     "url": "https://github.com/tiiuae/rclgo",
     "note": "ROS client library for Go."
    },
    {
     "name": "navigation",
     "url": "https://github.com/tiiuae/navigation",
     "note": "ROS2 3D navigation package."
    }
   ],
   "hf_highlights": [
    {
     "name": "tiiuae/falcon-7b",
     "url": "https://huggingface.co/tiiuae/falcon-7b",
     "kind": "model",
     "note": "Most downloaded Falcon model."
    },
    {
     "name": "tiiuae/falcon-7b-instruct",
     "url": "https://huggingface.co/tiiuae/falcon-7b-instruct",
     "kind": "model",
     "note": "Instruction-tuned 7B."
    },
    {
     "name": "tiiuae/falcon-40b",
     "url": "https://huggingface.co/tiiuae/falcon-40b",
     "kind": "model",
     "note": "40B base model."
    },
    {
     "name": "tiiuae/falcon-refinedweb",
     "url": "https://huggingface.co/datasets/tiiuae/falcon-refinedweb",
     "kind": "dataset",
     "note": "RefinedWeb dataset."
    }
   ],
   "papers": [
    {
     "title": "The Falcon Series of Open Language Models",
     "url": "https://arxiv.org/abs/2311.16867",
     "year": "2023"
    },
    {
     "title": "The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only",
     "url": "https://arxiv.org/abs/2306.01116",
     "year": "2023"
    },
    {
     "title": "Falcon Mamba: The First Competitive Attention-free 7B Language Model",
     "url": "https://arxiv.org/abs/2410.05355",
     "year": "2024"
    },
    {
     "title": "Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance",
     "url": "https://arxiv.org/abs/2507.22448",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Falcon-OCR",
     "url": "https://huggingface.co/tiiuae/Falcon-OCR",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Falcon-Perception",
     "url": "https://huggingface.co/tiiuae/Falcon-Perception",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "Falcon3-10B-Base-1.58bit-prequantized",
     "url": "https://huggingface.co/tiiuae/Falcon3-10B-Base-1.58bit-prequantized",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face and code on GitHub.",
   "key_people": [
    {
     "name": "Najwa Aaraj",
     "role": "Chief Executive Officer, TII",
     "url": "https://www.tii.ae/ai-and-digital-science",
     "note": null
    },
    {
     "name": "Hakim Hacid",
     "role": "Chief Researcher, AI Cross Center Unit",
     "url": "https://www.tii.ae/ai-and-digital-science",
     "note": null
    },
    {
     "name": "Yoelle Maarek",
     "role": "Chief Researcher, AI Cross Center Unit",
     "url": "https://www.tii.ae/ai-and-digital-science",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "tii.ae lists institute contact options; Falcon models and datasets are on Hugging Face.",
   "sources": [
    "https://huggingface.co/tiiuae",
    "https://www.tii.ae/"
   ],
   "confidence": "high",
   "region": "Middle East",
   "id": "technology-innovation-institute",
   "mlat": 24.47165,
   "mlon": 54.33881,
   "logo": true
  },
  {
   "orig_id": "alan-turing-institute",
   "name": "The Alan Turing Institute",
   "alias": null,
   "full_name": "The Alan Turing Institute",
   "type": "institute",
   "founded": 2015,
   "country": "United Kingdom",
   "city": "London",
   "lat": 51.5074,
   "lon": -0.1278,
   "website": "https://turing.ac.uk",
   "huggingface": "https://huggingface.co/alan-turing-institute",
   "github": "https://github.com/alan-turing-institute",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "UK national institute for data science and AI. It maintains a large body of open-source research software and releases some models and datasets.",
   "tagline": "The Alan Turing Institute is the UK's national institute for data science and AI, which maintains a large body of open research software.",
   "about": [
    "The Alan Turing Institute was founded in 2015 as the UK's national institute for data science, and AI was added to its remit in 2017. It is named after Alan Turing and is largely government funded.",
    "Its researchers and engineers publish many open repositories on GitHub. Since 2025 the institute has focused on defence and national security at the government's request.",
    "In May 2026 George Williamson, previously of HMGCC, was named chief executive."
   ],
   "mission": "The institute's stated direction is now to serve UK sovereign capability in AI, defence and security.",
   "notice": "The institute's direction was reoriented in 2025 toward defence and national security. Its open-source research software and datasets remain public.",
   "known_for": [
    "UK national AI institute",
    "Open research software",
    "Data science and AI research"
   ],
   "focus": [
    "research software",
    "datasets",
    "data science"
   ],
   "contributions": [
    {
     "title": "Institute founded",
     "note": "Established as the national institute for data science.",
     "year": "2015"
    },
    {
     "title": "AI added to remit",
     "note": "The institute's mission expands to AI.",
     "year": "2017"
    },
    {
     "title": "Defence and security focus",
     "note": "Strategy refocuses at the government's request.",
     "year": "2025"
    },
    {
     "title": "New chief executive",
     "note": "George Williamson appointed.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "CleverCSV",
     "url": "https://github.com/alan-turing-institute/CleverCSV",
     "note": "Python package for handling messy CSV files (1,344 stars)"
    },
    {
     "name": "ReadabiliPy",
     "url": "https://github.com/alan-turing-institute/ReadabiliPy",
     "note": "HTML content extractor in Python (361 stars)"
    },
    {
     "name": "AIrsenal",
     "url": "https://github.com/alan-turing-institute/AIrsenal",
     "note": "Machine learning Fantasy Premier League team (354 stars)"
    },
    {
     "name": "distinctipy",
     "url": "https://github.com/alan-turing-institute/distinctipy",
     "note": "Package for generating visually distinct colours (296 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "alan-turing-institute/turing-synthetic-radar-dataset",
     "url": "https://huggingface.co/datasets/alan-turing-institute/turing-synthetic-radar-dataset",
     "kind": "dataset",
     "note": "Synthetic radar dataset, the most downloaded Turing dataset."
    },
    {
     "name": "alan-turing-institute/mt5-large-finetuned-mnli-xtreme-xnli",
     "url": "https://huggingface.co/alan-turing-institute/mt5-large-finetuned-mnli-xtreme-xnli",
     "kind": "model",
     "note": "mT5 fine-tuned on multilingual inference data."
    },
    {
     "name": "alan-turing-institute/t0-s1.1-3B",
     "url": "https://huggingface.co/alan-turing-institute/t0-s1.1-3B",
     "kind": "model",
     "note": "3B T0 model variant."
    },
    {
     "name": "alan-turing-institute/t0-s1-14B",
     "url": "https://huggingface.co/alan-turing-institute/t0-s1-14B",
     "kind": "model",
     "note": "14B T0 model variant."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "CleverCSV",
     "url": "https://github.com/alan-turing-institute/CleverCSV",
     "note": "Package for handling messy CSV files"
    },
    {
     "name": "ReadabiliPy",
     "url": "https://github.com/alan-turing-institute/ReadabiliPy",
     "note": "HTML content extractor"
    },
    {
     "name": "Turing synthetic radar dataset",
     "url": "https://huggingface.co/datasets/alan-turing-institute/turing-synthetic-radar-dataset",
     "note": "Synthetic radar dataset"
    }
   ],
   "openness": "Large GitHub organisation with mostly MIT-licensed tools; Hugging Face releases are modest.",
   "key_people": [
    {
     "name": "George Williamson",
     "role": "Chief Executive",
     "url": "https://www.turing.ac.uk",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "The institute's website has a contact page; project-specific contact goes through each repository.",
   "sources": [
    "https://turing.ac.uk",
    "https://huggingface.co/alan-turing-institute",
    "https://github.com/alan-turing-institute"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "the-alan-turing-institute",
   "mlat": 51.46616,
   "mlon": -0.21089,
   "logo": true
  },
  {
   "orig_id": "tilde",
   "name": "Tilde",
   "alias": null,
   "full_name": "Tilde",
   "type": "startup",
   "founded": 1991,
   "country": "Latvia",
   "city": "Riga",
   "lat": 56.9496,
   "lon": 24.1052,
   "website": "https://tilde.ai",
   "huggingface": "https://huggingface.co/TildeAI",
   "github": "https://github.com/tilde-nlp",
   "x": "https://x.com/TildeCom",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/tilde-ai",
   "contact_email": "tilde@tilde.com",
   "summary": "Baltic language-technology company that developed TildeOpen, an open multilingual LLM for European languages. It also maintains machine translation and language tools.",
   "tagline": "Tilde is a Latvian language technology company that trained TildeOpen, an open 30B model covering 34 European languages.",
   "about": [
    "Tilde is a language technology company based in Latvia that builds machine translation and language tools. In 2024 it won the European Commission's Large AI Grand Challenge, which awarded 2 million GPU hours on the LUMI supercomputer.",
    "On that compute it trained TildeOpen LLM, a 30B dense decoder-only model on the Llama 3 architecture, trained on 2 trillion tokens for 34 European languages. It was released in October 2025 on Hugging Face under CC BY 4.0.",
    "A technical report presented at LREC 2026 describes the curriculum learning used to give lower-resource languages equitable representation."
   ],
   "mission": "Tilde aims to give smaller European languages an open model that serves them as well as large ones.",
   "notice": null,
   "known_for": [
    "TildeOpen 30B",
    "Equitable European language coverage",
    "Machine translation"
   ],
   "focus": [
    "language models",
    "machine translation",
    "Baltic languages",
    "multilingual"
   ],
   "contributions": [
    {
     "title": "Large AI Grand Challenge win",
     "note": "Awarded 2 million GPU hours on LUMI.",
     "year": "2024"
    },
    {
     "title": "TildeOpen LLM release",
     "note": "30B open model on Hugging Face, 15 October.",
     "year": "2025"
    },
    {
     "title": "LREC 2026 paper",
     "note": "Technical report on curriculum learning for equitable representation.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "terminology_translation",
     "url": "https://github.com/tilde-nlp/terminology_translation",
     "note": "Open-source repository (5 stars)"
    },
    {
     "name": "website-translator",
     "url": "https://github.com/tilde-nlp/website-translator",
     "note": "Website translation JS library for website localization (5 stars)"
    },
    {
     "name": "c-eval",
     "url": "https://github.com/tilde-nlp/c-eval",
     "note": "Parallel corpora cleaning and evaluation tool (4 stars)"
    },
    {
     "name": "et-mt-tools",
     "url": "https://github.com/tilde-nlp/et-mt-tools",
     "note": "This repository contains guidelines and tools for SMT system developement researched within the project \"EKT63 (4 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "TildeAI/TildeOpen-30b",
     "url": "https://huggingface.co/TildeAI/TildeOpen-30b",
     "kind": "model",
     "note": "Model: 2,243 downloads, 172 likes."
    },
    {
     "name": "TildeAI/TildeOpen-30b-64k",
     "url": "https://huggingface.co/TildeAI/TildeOpen-30b-64k",
     "kind": "model",
     "note": "Model: 594 downloads, 3 likes."
    },
    {
     "name": "TildeAI/TildeOpen-MT-Euro-mix",
     "url": "https://huggingface.co/datasets/TildeAI/TildeOpen-MT-Euro-mix",
     "kind": "dataset",
     "note": "Dataset: 65 downloads, 1 likes."
    },
    {
     "name": "TildeAI/TildeOpen8B-64k-wmt26-compression-task-cs-de-fp4a16",
     "url": "https://huggingface.co/TildeAI/TildeOpen8B-64k-wmt26-compression-task-cs-de-fp4a16",
     "kind": "model",
     "note": "Model: 17 downloads, 0 likes."
    }
   ],
   "papers": [
    {
     "title": "TildeOpen LLM: Leveraging Curriculum Learning to Achieve Equitable Language Representation",
     "url": "https://arxiv.org/abs/2603.08182",
     "year": "2026"
    }
   ],
   "notable_releases": [
    {
     "name": "TildeOpen-30b",
     "url": "https://huggingface.co/TildeAI/TildeOpen-30b",
     "note": "Open multilingual 30B model focused on European languages"
    },
    {
     "name": "TildeOpen-30b-64k",
     "url": "https://huggingface.co/TildeAI/TildeOpen-30b-64k",
     "note": "Long-context variant"
    },
    {
     "name": "TildeOpen-MT-Euro-mix",
     "url": "https://huggingface.co/datasets/TildeAI/TildeOpen-MT-Euro-mix",
     "note": "Translation data mix"
    }
   ],
   "openness": "TildeOpen weights are public on Hugging Face; a number of tools are on GitHub under open licences.",
   "key_people": [
    {
     "name": "Mārcis Pinnis",
     "role": "Senior author, TildeOpen",
     "url": "https://arxiv.org/abs/2603.08182",
     "note": "Last author of the TildeOpen report."
    },
    {
     "name": "Toms Bergmanis",
     "role": "Lead author, TildeOpen",
     "url": "https://arxiv.org/abs/2603.08182",
     "note": "First author of the TildeOpen report."
    },
    {
     "name": "Artūrs Vasiļevskis",
     "role": "Chief executive quoted in launch coverage",
     "url": "https://tilde.ai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on the website.",
   "sources": [
    "https://tilde.ai",
    "https://huggingface.co/TildeAI",
    "https://github.com/tilde-nlp"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "tilde",
   "mlat": 56.98496,
   "mlon": 24.1052,
   "logo": true
  },
  {
   "orig_id": "tng",
   "name": "TNG Technology Consulting",
   "alias": null,
   "full_name": "TNG Technology Consulting",
   "type": "startup",
   "founded": 2001,
   "country": "Germany",
   "city": "Munich",
   "lat": 48.1351,
   "lon": 11.582,
   "website": "https://tngtech.com",
   "huggingface": "https://huggingface.co/tngtech",
   "github": null,
   "x": "https://x.com/tngtech",
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/tng-technology-consulting",
   "contact_email": "info@tngtech.com",
   "summary": "German technology consultancy that publishes open-weight model merges such as the DeepSeek R1T Chimera models on Hugging Face.",
   "tagline": "TNG is a German technology consultancy that publishes open-weight DeepSeek merges such as the R1T Chimera models.",
   "about": [
    "TNG Technology Consulting is a Bavarian consultancy founded in 2001 with several hundred employees. Its AI research team publishes model work on Hugging Face under tngtech.",
    "DeepSeek-R1T-Chimera, released on 27 April 2025, is constructed from parts of DeepSeek-R1 and DeepSeek-V3 mixture-of-experts models rather than fine-tuning. DeepSeek-TNG R1T2 Chimera followed on 3 July 2025 under the MIT license."
   ],
   "mission": "TNG presents its model work as research on assembling new models from existing open parents.",
   "notice": null,
   "known_for": [
    "DeepSeek R1T Chimera",
    "R1T2 Chimera model merges",
    "Open-weight model assembly"
   ],
   "focus": [
    "language models",
    "model merging"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "TNG is established.",
     "year": "2001"
    },
    {
     "title": "DeepSeek-R1T-Chimera",
     "note": "Released 27 April.",
     "year": "2025"
    },
    {
     "title": "DeepSeek-TNG R1T2 Chimera",
     "note": "Released 3 July with MIT-licensed weights.",
     "year": "2025"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "tngtech/DeepSeek-TNG-R1T2-Chimera",
     "url": "https://huggingface.co/tngtech/DeepSeek-TNG-R1T2-Chimera",
     "kind": "model",
     "note": "Merged reasoning model with 288 likes."
    },
    {
     "name": "tngtech/DeepSeek-R1T-Chimera",
     "url": "https://huggingface.co/tngtech/DeepSeek-R1T-Chimera",
     "kind": "model",
     "note": "First Chimera model."
    },
    {
     "name": "tngtech/olmOCR-7B-faithful",
     "url": "https://huggingface.co/tngtech/olmOCR-7B-faithful",
     "kind": "model",
     "note": "Fine-tuned OCR model."
    },
    {
     "name": "tngtech/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-GGUF",
     "url": "https://huggingface.co/tngtech/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-GGUF",
     "kind": "model",
     "note": "Quantised GGUF of a Nemotron model."
    }
   ],
   "papers": [
    {
     "title": "Assembly of Experts: Linear-time construction of the Chimera LLM variants with emergent and adaptable behaviors",
     "url": "https://arxiv.org/abs/2506.14794",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "DeepSeek-TNG-R1T2-Chimera",
     "url": "https://huggingface.co/tngtech/DeepSeek-TNG-R1T2-Chimera",
     "note": "Merged reasoning model"
    },
    {
     "name": "DeepSeek-R1T-Chimera",
     "url": "https://huggingface.co/tngtech/DeepSeek-R1T-Chimera",
     "note": "First Chimera model"
    },
    {
     "name": "olmOCR-7B-faithful",
     "url": "https://huggingface.co/tngtech/olmOCR-7B-faithful",
     "note": "OCR model"
    }
   ],
   "openness": "Weights are public on Hugging Face; models derive from third-party bases and carry their licences.",
   "key_people": [
    {
     "name": "Henrik Klagges",
     "role": "Lead author, Assembly of Experts",
     "url": "https://arxiv.org/abs/2506.14794",
     "note": "First author of the Chimera paper."
    },
    {
     "name": "Robert Dahlke",
     "role": "Co-author",
     "url": "https://arxiv.org/abs/2506.14794",
     "note": "Co-author of the Chimera paper."
    },
    {
     "name": "Fabian Klemm",
     "role": "Co-author",
     "url": "https://arxiv.org/abs/2506.14794",
     "note": "Co-author of the Chimera paper."
    }
   ],
   "discord_note": null,
   "contact_note": "The company's contact page prints a general address; the Chimera work is documented on its Hugging Face page.",
   "sources": [
    "https://tngtech.com",
    "https://huggingface.co/tngtech"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "tng-technology-consulting",
   "mlat": 48.11783,
   "mlon": 11.50234,
   "logo": true
  },
  {
   "orig_id": "together-ai",
   "name": "Together AI",
   "alias": null,
   "full_name": "Together AI",
   "type": "startup",
   "founded": 2022,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.together.ai",
   "huggingface": "https://huggingface.co/togethercomputer",
   "github": "https://github.com/togethercomputer",
   "x": null,
   "discord": "https://discord.gg/9Rk6sSeWEG",
   "linkedin": null,
   "contact_email": null,
   "summary": "AI cloud company whose research arm releases open datasets, models and methods, including RedPajama data and models and Mixture-of-Agents.",
   "tagline": "Together AI is an AI cloud company whose research arm released RedPajama and advanced open work such as FlashAttention-2 and Mixture-of-Agents.",
   "about": [
    "Together AI was founded by Vipul Ved Prakash, Ce Zhang, Chris Ré and Percy Liang, with Tri Dao joining as chief scientist in July 2023. It combines a commercial inference cloud with an open research group.",
    "In April 2023 it released RedPajama, a 1.2 trillion token pre-training dataset reproducing the LLaMA recipe, and later the 30 trillion token RedPajama-V2. It announced FlashAttention-2 in July 2023.",
    "Its research outputs are open, but the company's main business is a paid platform."
   ],
   "mission": "Together aims to provide an open, high-performance platform for building with open-source models.",
   "notice": null,
   "known_for": [
    "RedPajama datasets",
    "FlashAttention-2",
    "Open model cloud"
   ],
   "focus": [
    "datasets",
    "language models",
    "training infrastructure",
    "inference"
   ],
   "contributions": [
    {
     "title": "RedPajama dataset",
     "note": "1.2 trillion token open dataset released 17 April.",
     "year": "2023"
    },
    {
     "title": "FlashAttention-2",
     "note": "Announced 17 July with Tri Dao as chief scientist.",
     "year": "2023"
    },
    {
     "title": "RedPajama-V2",
     "note": "A 30 trillion token open dataset.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "OpenChatKit",
     "url": "https://github.com/togethercomputer/OpenChatKit",
     "note": "Open chatbot toolkit; most-starred repo."
    },
    {
     "name": "RedPajama-Data",
     "url": "https://github.com/togethercomputer/RedPajama-Data",
     "note": "Code for preparing large-scale pretraining data."
    },
    {
     "name": "MoA",
     "url": "https://github.com/togethercomputer/MoA",
     "note": "Together Mixture-Of-Agents."
    },
    {
     "name": "together-cookbook",
     "url": "https://github.com/togethercomputer/together-cookbook",
     "note": "Recipes for open models on Together."
    }
   ],
   "hf_highlights": [
    {
     "name": "LLaMA-2-7B-32K",
     "url": "https://huggingface.co/togethercomputer/LLaMA-2-7B-32K",
     "kind": "model",
     "note": "Llama 2 extended to 32K context."
    },
    {
     "name": "RedPajama-Data-V2",
     "url": "https://huggingface.co/datasets/togethercomputer/RedPajama-Data-V2",
     "kind": "dataset",
     "note": "Large web corpus with quality annotations."
    },
    {
     "name": "RedPajama-Data-1T",
     "url": "https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T",
     "kind": "dataset",
     "note": "1T-token open replication of LLaMA data."
    }
   ],
   "papers": [
    {
     "title": "RedPajama: an Open Dataset for Training Large Language Models",
     "url": "https://arxiv.org/abs/2411.12372",
     "year": "2024"
    },
    {
     "title": "Mixture-of-Agents Enhances Large Language Model Capabilities",
     "url": "https://arxiv.org/abs/2406.04692",
     "year": "2024"
    },
    {
     "title": "FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision",
     "url": "https://arxiv.org/abs/2407.08608",
     "year": "2024"
    },
    {
     "title": "Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads",
     "url": "https://arxiv.org/abs/2401.10774",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "RedPajama-Data",
     "url": "https://github.com/togethercomputer/RedPajama-Data",
     "note": "Open pretraining dataset tooling"
    },
    {
     "name": "MoA",
     "url": "https://github.com/togethercomputer/MoA",
     "note": "Mixture-of-Agents code"
    },
    {
     "name": "OpenChatKit",
     "url": "https://github.com/togethercomputer/OpenChatKit",
     "note": "Open chat model toolkit"
    }
   ],
   "openness": "Open datasets, weights and code; commercial cloud business alongside.",
   "key_people": [
    {
     "name": "Vipul Ved Prakash",
     "role": "Co-founder and CEO",
     "url": "https://www.together.ai",
     "note": null
    },
    {
     "name": "Ce Zhang",
     "role": "Co-founder",
     "url": "https://www.together.ai",
     "note": null
    },
    {
     "name": "Chris Ré",
     "role": "Co-founder",
     "url": "https://www.together.ai",
     "note": null
    },
    {
     "name": "Percy Liang",
     "role": "Co-founder",
     "url": "https://cs.stanford.edu/~pliang/",
     "note": null
    },
    {
     "name": "Tri Dao",
     "role": "Chief Scientist",
     "url": "https://tridao.me",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.together.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.together.ai",
    "https://huggingface.co/togethercomputer",
    "https://github.com/togethercomputer"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "together-ai",
   "mlat": 37.69721,
   "mlon": -122.53511,
   "logo": true
  },
  {
   "orig_id": "transluce",
   "name": "Transluce",
   "alias": null,
   "full_name": "Transluce",
   "type": "nonprofit",
   "founded": 2024,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://transluce.org",
   "huggingface": "https://huggingface.co/Transluce",
   "github": "https://github.com/TransluceAI",
   "x": "https://x.com/TransluceAI",
   "discord": null,
   "linkedin": null,
   "contact_email": "info@transluce.org",
   "summary": "Independent nonprofit AI research lab building open tools to understand and steer AI systems, including interpretability toolkits and agentic investigation tools.",
   "tagline": "Transluce is a nonprofit AI research lab building open tools to understand and oversee AI systems.",
   "about": [
    "Transluce was co-founded in 2024 by Jacob Steinhardt and Sarah Schwettmann, with Steinhardt on leave from UC Berkeley and Schwettmann previously at MIT CSAIL. It describes itself as a nonprofit that builds open, scalable tools for understanding frontier AI.",
    "It launched publicly on 23 October 2024 with neuron-description tools, Monitor and investigator agents. It later released Docent, which analyzes agent transcripts to find bugs and unexpected behaviour and is used by more than 25 organizations."
   ],
   "mission": "Transluce aims to enable scalable, democratic oversight of AI through open tooling.",
   "notice": null,
   "known_for": [
    "Docent agent transcript analysis",
    "Interpretability tooling",
    "Nonprofit AI oversight research"
   ],
   "focus": [
    "interpretability",
    "evaluation",
    "safety",
    "tooling"
   ],
   "contributions": [
    {
     "title": "Lab co-founded",
     "note": "Steinhardt and Schwettmann start Transluce.",
     "year": "2024"
    },
    {
     "title": "Public launch",
     "note": "Tools and investigator agents unveiled 23 October.",
     "year": "2024"
    },
    {
     "title": "Docent",
     "note": "Released for analyzing AI agent transcripts.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "observatory",
     "url": "https://github.com/TransluceAI/observatory",
     "note": "Describe model features and intervene on them"
    },
    {
     "name": "docent",
     "url": "https://github.com/TransluceAI/docent",
     "note": "Agent transcript analysis"
    },
    {
     "name": "circuits",
     "url": "https://github.com/TransluceAI/circuits",
     "note": "MLP neuron-level circuit tracing"
    },
    {
     "name": "introspective-interp",
     "url": "https://github.com/TransluceAI/introspective-interp",
     "note": "Models that explain their own computations"
    }
   ],
   "hf_highlights": [
    {
     "name": "llama_8b_explainer",
     "url": "https://huggingface.co/Transluce/llama_8b_explainer",
     "kind": "model",
     "note": "Feature explainer model"
    },
    {
     "name": "llama_8b_simulator",
     "url": "https://huggingface.co/Transluce/llama_8b_simulator",
     "kind": "model",
     "note": "Feature simulator model"
    },
    {
     "name": "WeirdChat",
     "url": "https://huggingface.co/datasets/Transluce/WeirdChat",
     "kind": "dataset",
     "note": "Chat dataset"
    }
   ],
   "papers": [
    {
     "title": "Eliciting Language Model Behaviors with Investigator Agents",
     "url": "https://arxiv.org/abs/2502.01236",
     "year": "2025"
    },
    {
     "title": "Training Language Models to Explain Their Own Computations",
     "url": "https://arxiv.org/abs/2511.08579",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "observatory",
     "url": "https://github.com/TransluceAI/observatory",
     "note": "Describe model features and intervene on them"
    },
    {
     "name": "docent",
     "url": "https://github.com/TransluceAI/docent",
     "note": "Agent transcript analysis"
    },
    {
     "name": "circuits",
     "url": "https://github.com/TransluceAI/circuits",
     "note": "MLP neuron-level circuit tracing"
    }
   ],
   "openness": "MIT-licensed code on GitHub; explainer models and datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Jacob Steinhardt",
     "role": "Co-founder and CEO",
     "url": "https://jsteinhardt.stat.berkeley.edu/",
     "note": "On leave from UC Berkeley."
    },
    {
     "name": "Sarah Schwettmann",
     "role": "Co-founder and Chief Scientist",
     "url": "https://transluce.org/about",
     "note": "Formerly at MIT CSAIL."
    },
    {
     "name": "Kevin Meng",
     "role": "Member of technical staff",
     "url": "https://mengk.me/",
     "note": "Researcher on interpretability tooling."
    },
    {
     "name": "Robin Jia",
     "role": "Scientific advisor",
     "url": "https://robinjia.github.io/",
     "note": "Listed on the team page."
    }
   ],
   "discord_note": null,
   "contact_note": "Email printed on website; also via site.",
   "sources": [
    "https://transluce.org",
    "https://github.com/TransluceAI",
    "https://huggingface.co/Transluce"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "transluce",
   "mlat": 37.67937,
   "mlon": -122.51128,
   "logo": true
  },
  {
   "orig_id": "trillion-labs",
   "name": "Trillion Labs",
   "alias": null,
   "full_name": "Trillion Labs",
   "type": "startup",
   "founded": 2024,
   "country": "South Korea",
   "city": "Seoul",
   "lat": 37.57,
   "lon": 126.98,
   "website": "https://trillionlabs.co",
   "huggingface": "https://huggingface.co/trillionlabs",
   "github": "https://github.com/trillion-labs",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Korean foundation model startup publishing models such as Zoro and GravityOCR-KO.",
   "tagline": "Trillion Labs is a Seoul startup that trains Korean and English foundation models from scratch and releases open weights.",
   "about": [
    "Trillion Labs was founded in Seoul in 2024 by Jaemin Shin, formerly a research scientist on Naver's HyperCLOVA X. It raised pre-seed funding of 5.8 million dollars in September 2024.",
    "It open-sourced the preview Trillion-7B model in March 2025 and Tri-21B in July 2025, a model trained from scratch that runs on a single GPU. Tri 21B Think, adding reinforcement learning for reasoning, followed in February 2026."
   ],
   "mission": "Trillion Labs aims to build efficient Korean-capable models through from-scratch pretraining.",
   "notice": null,
   "known_for": [
    "Tri model series",
    "Korean from-scratch pretraining",
    "Tri-21B"
   ],
   "focus": [
    "language models",
    "Korean",
    "OCR"
   ],
   "contributions": [
    {
     "title": "Pre-seed round",
     "note": "5.8 million dollars raised in September.",
     "year": "2024"
    },
    {
     "title": "Trillion-7B-preview",
     "note": "Open preview model released.",
     "year": "2025"
    },
    {
     "title": "Tri-21B",
     "note": "Open from-scratch 21B model released in July.",
     "year": "2025"
    },
    {
     "title": "Tri 21B Think",
     "note": "Reasoning variant unveiled.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "gWorld",
     "url": "https://github.com/trillion-labs/gWorld",
     "note": "Open-weight GUI world model for mobile agents."
    },
    {
     "name": "trida-stack",
     "url": "https://github.com/trillion-labs/trida-stack",
     "note": "Convert and serve block-diffusion LLMs."
    },
    {
     "name": "claude-code-orchestrator",
     "url": "https://github.com/trillion-labs/claude-code-orchestrator",
     "note": "Dashboard for managing multiple Claude Code sessions."
    },
    {
     "name": "nemoslides",
     "url": "https://github.com/trillion-labs/nemoslides",
     "note": "Open-weight slide generation fine-tuned on Nemotron."
    }
   ],
   "hf_highlights": [
    {
     "name": "trillionlabs/Tri-7B",
     "url": "https://huggingface.co/trillionlabs/Tri-7B",
     "kind": "model",
     "note": "Tri 7B base/instruct model."
    },
    {
     "name": "trillionlabs/TxGravity-30B-A5B-Open",
     "url": "https://huggingface.co/trillionlabs/TxGravity-30B-A5B-Open",
     "kind": "model",
     "note": "Open 30B MoE model."
    },
    {
     "name": "trillionlabs/Tri-1.8B-Translation",
     "url": "https://huggingface.co/trillionlabs/Tri-1.8B-Translation",
     "kind": "model",
     "note": "Small translation model."
    },
    {
     "name": "trillionlabs/TheBioCollection",
     "url": "https://huggingface.co/datasets/trillionlabs/TheBioCollection",
     "kind": "dataset",
     "note": "Bio-domain dataset."
    }
   ],
   "papers": [
    {
     "title": "Trillion 7B Technical Report",
     "url": "https://arxiv.org/abs/2504.15431",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "Zoro-1",
     "url": "https://huggingface.co/trillionlabs/Zoro-1",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "GravityOCR-KO",
     "url": "https://huggingface.co/trillionlabs/GravityOCR-KO",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face.",
   "key_people": [
    {
     "name": "Jamin Shin",
     "role": "Founder and CEO",
     "url": "https://arxiv.org/abs/2504.15431",
     "note": "Last author of the Trillion 7B report; formerly at Naver HyperCLOVA X."
    },
    {
     "name": "Sungjun Han",
     "role": "Researcher, first author",
     "url": "https://arxiv.org/abs/2504.15431",
     "note": "First author of the Trillion 7B report."
    },
    {
     "name": "Juyoung Suk",
     "role": "Researcher",
     "url": "https://arxiv.org/abs/2504.15431",
     "note": "Co-author of the Trillion 7B report."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact options are given on trillionlabs.co; their code and models are on GitHub and Hugging Face.",
   "sources": [
    "https://huggingface.co/trillionlabs",
    "https://trillionlabs.co"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "trillion-labs",
   "mlat": 37.59165,
   "mlon": 127.02731,
   "logo": true
  },
  {
   "orig_id": "turing",
   "name": "Turing Inc.",
   "alias": null,
   "full_name": "Turing Inc.",
   "type": "startup",
   "founded": 2021,
   "country": "Japan",
   "city": "Kashiwa",
   "lat": 35.86,
   "lon": 139.97,
   "website": "https://tur.ing/en",
   "huggingface": "https://huggingface.co/turing-motors",
   "github": "https://github.com/turingmotors",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Autonomous driving company that releases the Heron vision-language models and related tooling.",
   "tagline": "Turing is a Tokyo autonomous-driving company that releases the Heron vision-language models and datasets.",
   "about": [
    "Turing was founded in 2021 in Tokyo with the goal of fully autonomous vehicles by 2030. Issei Yamamoto is chief executive and Shunsuke Aoki co-founder and CTO.",
    "Its Heron family of vision-language models, supported by Japan's GENIAC program, includes the Heron-NVILA-Lite 15B, 2B and 1B models. In March 2026 it reported Japan's first real-time public-road driving controlled by a vision-language-action model."
   ],
   "mission": "Turing aims to deliver end-to-end AI driving for fully autonomous cars.",
   "notice": null,
   "known_for": [
    "Heron vision-language models",
    "End-to-end autonomous driving",
    "CoVLA dataset"
   ],
   "focus": [
    "vision-language",
    "autonomous driving"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Turing starts with the aim of fully autonomous vehicles.",
     "year": "2021"
    },
    {
     "title": "Heron and CoVLA",
     "note": "Heron VLMs and the CoVLA driving dataset released.",
     "year": "2024"
    },
    {
     "title": "Heron-NVILA-Lite",
     "note": "15B, 2B and 1B models made public under GENIAC.",
     "year": "2025"
    },
    {
     "title": "VLA driving on public roads",
     "note": "First real-time VLA control reported in Japan.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "swan",
     "url": "https://github.com/turingmotors/swan",
     "note": "LM inference on FPGAs."
    },
    {
     "name": "heron",
     "url": "https://github.com/turingmotors/heron",
     "note": "Vision and language model library."
    },
    {
     "name": "One-D-Piece",
     "url": "https://github.com/turingmotors/One-D-Piece",
     "note": "Image tokenizer with quality-controllable compression."
    },
    {
     "name": "NuScenes-MQA",
     "url": "https://github.com/turingmotors/NuScenes-MQA",
     "note": "Driving QA dataset."
    }
   ],
   "hf_highlights": [
    {
     "name": "Heron-NVILA-Lite-2B",
     "url": "https://huggingface.co/turing-motors/Heron-NVILA-Lite-2B",
     "kind": "model",
     "note": "Small Heron VLM."
    },
    {
     "name": "Heron-NVILA-Lite-1B",
     "url": "https://huggingface.co/turing-motors/Heron-NVILA-Lite-1B",
     "kind": "model",
     "note": "1B VLM."
    },
    {
     "name": "Cauldron-JA",
     "url": "https://huggingface.co/datasets/turing-motors/Cauldron-JA",
     "kind": "dataset",
     "note": "Japanese VL dataset."
    },
    {
     "name": "CoVLA-Dataset",
     "url": "https://huggingface.co/datasets/turing-motors/CoVLA-Dataset",
     "kind": "dataset",
     "note": "Driving dataset."
    }
   ],
   "papers": [
    {
     "title": "CoVLA: Comprehensive Vision-Language-Action Dataset for Autonomous Driving",
     "url": "https://arxiv.org/abs/2408.10845",
     "year": "2024"
    },
    {
     "title": "Heron-Bench: A Benchmark for Evaluating Vision Language Models in Japanese",
     "url": "https://arxiv.org/abs/2404.07824",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Heron-NVILA-Lite-33B-hf",
     "url": "https://huggingface.co/turing-motors/Heron-NVILA-Lite-33B-hf",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "drive_titok_s_640",
     "url": "https://huggingface.co/turing-motors/drive_titok_s_640",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Weights on Hugging Face, code on GitHub.",
   "key_people": [
    {
     "name": "Issei Yamamoto",
     "role": "CEO",
     "url": "https://tur.ing/en",
     "note": null
    },
    {
     "name": "Shunsuke Aoki",
     "role": "Co-founder and CTO",
     "url": "https://tur.ing/en",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via its website.",
   "sources": [
    "https://huggingface.co/turing-motors",
    "https://tur.ing/en"
   ],
   "confidence": "medium",
   "region": "Asia",
   "id": "turing-inc",
   "logo": true
  },
  {
   "orig_id": "turkunlp",
   "name": "TurkuNLP",
   "alias": null,
   "full_name": "TurkuNLP",
   "type": "academic",
   "founded": null,
   "country": "Finland",
   "city": "Turku",
   "lat": 60.4518,
   "lon": 22.2666,
   "website": "https://turkunlp.org",
   "huggingface": "https://huggingface.co/TurkuNLP",
   "github": "https://github.com/TurkuNLP",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "NLP research group at the University of Turku, known for FinBERT and the Poro models. It has released many Finnish and multilingual models, parsers and datasets.",
   "tagline": "TurkuNLP is the University of Turku's language technology group, known for FinBERT and, with Silo AI, the Poro and Viking models.",
   "about": [
    "TurkuNLP is a group of about 30 researchers at the University of Turku, ranging from research assistants to professors. Its main topics are syntactic and semantic analysis of Finnish, large language models for Finnish and other languages, biomedical and clinical text mining, and the analysis of internet registers and genres.",
    "The group's resources include FinBERT, WikiBERT, FinGPT and the Turku Dependency Treebank. With Silo AI's SiloGen it built the Poro and Viking models on the LUMI supercomputer; Viking, released in 2024 in 7B, 13B and 33B sizes, covers the Nordic languages, English and code under Apache 2.0.",
    "TurkuNLP co-hosts an open monthly seminar on NLP and AI with Omni Language AI Research, and publishes its code and models on GitHub and Hugging Face."
   ],
   "mission": "The group aims to produce open language resources and models for Finnish and the Nordic languages.",
   "notice": null,
   "known_for": [
    "FinBERT",
    "Poro and Viking Nordic models",
    "Universal Dependencies treebanks"
   ],
   "focus": [
    "language models",
    "Finnish",
    "datasets",
    "parsing"
   ],
   "contributions": [
    {
     "title": "FinBERT",
     "note": "A Finnish BERT model released.",
     "year": "2019"
    },
    {
     "title": "Poro",
     "note": "Open model for Finnish trained on LUMI with Silo AI.",
     "year": "2023"
    },
    {
     "title": "Viking 7B, 13B and 33B",
     "note": "Nordic model family released.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "Turku-neural-parser-pipeline",
     "url": "https://github.com/TurkuNLP/Turku-neural-parser-pipeline",
     "note": "A neural parsing pipeline for segmentation, morphological tagging, dependency parsing and lemmatization with p (116 stars)"
    },
    {
     "name": "FinBERT",
     "url": "https://github.com/TurkuNLP/FinBERT",
     "note": "BERT model trained from scratch on Finnish (100 stars)"
    },
    {
     "name": "Finnish-dep-parser",
     "url": "https://github.com/TurkuNLP/Finnish-dep-parser",
     "note": "The Finnish dependency parsing pipeline being developed by the Turku NLP group. Documentation: (51 stars)"
    },
    {
     "name": "wikibert",
     "url": "https://github.com/TurkuNLP/wikibert",
     "note": "BERT models for many languages created from Wikipedia texts (33 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "TurkuNLP/sbert-cased-finnish-paraphrase",
     "url": "https://huggingface.co/TurkuNLP/sbert-cased-finnish-paraphrase",
     "kind": "model",
     "note": "Model: 10,175 downloads, 4 likes."
    },
    {
     "name": "TurkuNLP/bert-base-finnish-cased-v1",
     "url": "https://huggingface.co/TurkuNLP/bert-base-finnish-cased-v1",
     "kind": "model",
     "note": "Model: 4,472 downloads, 10 likes."
    },
    {
     "name": "TurkuNLP/register_oscar",
     "url": "https://huggingface.co/datasets/TurkuNLP/register_oscar",
     "kind": "dataset",
     "note": "Dataset: 4,215 downloads, 5 likes."
    },
    {
     "name": "TurkuNLP/finnish-modernbert-base",
     "url": "https://huggingface.co/TurkuNLP/finnish-modernbert-base",
     "kind": "model",
     "note": "Model: 1,060 downloads, 0 likes."
    }
   ],
   "papers": [
    {
     "title": "Multilingual is not enough: BERT for Finnish",
     "url": "https://arxiv.org/abs/1912.07076",
     "year": "2019"
    },
    {
     "title": "Poro 34B and the Blessing of Multilinguality",
     "url": "https://arxiv.org/abs/2404.01856",
     "year": "2024"
    },
    {
     "title": "FinGPT: Large Generative Models for a Small Language",
     "url": "https://arxiv.org/abs/2311.05640",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "FinBERT",
     "url": "https://github.com/TurkuNLP/FinBERT",
     "note": "BERT model trained from scratch on Finnish"
    },
    {
     "name": "Turku neural parser pipeline",
     "url": "https://github.com/TurkuNLP/Turku-neural-parser-pipeline",
     "note": "Dependency parsing pipeline"
    },
    {
     "name": "gpt3-finnish-13B",
     "url": "https://huggingface.co/TurkuNLP/gpt3-finnish-13B",
     "note": "Finnish GPT-style model"
    },
    {
     "name": "FinerWeb-10BT",
     "url": "https://huggingface.co/datasets/TurkuNLP/finerweb-10bt",
     "note": "Filtered Finnish web data"
    }
   ],
   "openness": "Models, datasets and tools are public on Hugging Face and GitHub; licences vary by project.",
   "key_people": [
    {
     "name": "Filip Ginter",
     "role": "Professor",
     "url": "https://fginter.github.io/",
     "note": "Co-leads the group's Finnish language-model and parsing work."
    },
    {
     "name": "Sampo Pyysalo",
     "role": "Senior Research Fellow",
     "url": "https://www.utu.fi/en/people/sampo-pyysalo",
     "note": "Researcher in language models and biomedical text mining."
    },
    {
     "name": "Veronika Laippala",
     "role": "Professor",
     "url": "https://www.utu.fi/en/people/veronika-laippala",
     "note": "Leads work on web registers and genres."
    },
    {
     "name": "Tapio Salakoski",
     "role": "Professor",
     "url": "https://www.utu.fi/en/people/tapio-salakoski",
     "note": "Senior professor of the group."
    },
    {
     "name": "Jenna Kanerva",
     "role": "University Lecturer",
     "url": "https://github.com/jmnybl",
     "note": "Researcher on Finnish language technology."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://turkunlp.org",
    "https://huggingface.co/TurkuNLP",
    "https://github.com/TurkuNLP",
    "https://arxiv.org/abs/1912.07076"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "turkunlp",
   "logo": true
  },
  {
   "orig_id": "ukp-lab",
   "name": "UKP Lab",
   "alias": "TU Darmstadt",
   "full_name": "UKP Lab (TU Darmstadt)",
   "type": "academic",
   "founded": 2009,
   "country": "Germany",
   "city": "Darmstadt",
   "lat": 49.8728,
   "lon": 8.6512,
   "website": "https://www.ukp.tu-darmstadt.de",
   "huggingface": "https://huggingface.co/UKPLab",
   "github": "https://github.com/UKPLab",
   "x": null,
   "discord": null,
   "linkedin": "https://www.linkedin.com/company/ukplab",
   "contact_email": null,
   "summary": "NLP research lab at TU Darmstadt that created Sentence-BERT and publishes open code, models and datasets.",
   "tagline": "The Ubiquitous Knowledge Processing Lab at TU Darmstadt created Sentence-BERT and publishes open NLP tools.",
   "about": [
    "The Ubiquitous Knowledge Processing Lab belongs to the Department of Computer Science at TU Darmstadt and was founded by Iryna Gurevych. Its early work covered information retrieval, question answering and the structured use of wikis, including the Java Wikipedia Library and Java Wiktionary Library.",
    "It created Sentence-BERT, the basis of the Sentence Transformers library, with Nils Reimers in 2019, and develops AdapterHub. In October 2025 maintenance of Sentence Transformers passed to Hugging Face under Apache 2.0."
   ],
   "mission": "The lab works on open, reusable language technology and knowledge processing methods.",
   "notice": null,
   "known_for": [
    "Sentence-BERT",
    "Sentence Transformers",
    "AdapterHub"
   ],
   "focus": [
    "NLP",
    "embeddings",
    "datasets"
   ],
   "contributions": [
    {
     "title": "Lab founded",
     "note": "Iryna Gurevych founds UKP Lab.",
     "year": "2009"
    },
    {
     "title": "Sentence-BERT",
     "note": "The library emerges from lab research.",
     "year": "2019"
    },
    {
     "title": "AdapterHub",
     "note": "Framework paper published.",
     "year": "2020"
    },
    {
     "title": "Transfer to Hugging Face",
     "note": "Sentence Transformers maintenance moves to Hugging Face.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "EasyNMT",
     "url": "https://github.com/UKPLab/EasyNMT",
     "note": "Easy-to-use neural machine translation for 100+ languages (1,260 stars)"
    },
    {
     "name": "emnlp2017-bilstm-cnn-crf",
     "url": "https://github.com/UKPLab/emnlp2017-bilstm-cnn-crf",
     "note": "BiLSTM-CNN-CRF sequence tagging (834 stars)"
    },
    {
     "name": "gpl",
     "url": "https://github.com/UKPLab/gpl",
     "note": "Unsupervised domain adaptation for dense retrieval (344 stars)"
    },
    {
     "name": "deeplearning4nlp-tutorial",
     "url": "https://github.com/UKPLab/deeplearning4nlp-tutorial",
     "note": "Hands-on tutorial on deep learning for NLP (649 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "UKPLab/GritHopper-7B",
     "url": "https://huggingface.co/UKPLab/GritHopper-7B",
     "kind": "model",
     "note": "7B multi-hop retrieval and reasoning model."
    },
    {
     "name": "UKPLab/dapr",
     "url": "https://huggingface.co/datasets/UKPLab/dapr",
     "kind": "dataset",
     "note": "Document-aware passage retrieval benchmark."
    },
    {
     "name": "UKPLab/misviz",
     "url": "https://huggingface.co/datasets/UKPLab/misviz",
     "kind": "dataset",
     "note": "Misleading visualisation dataset."
    },
    {
     "name": "UKPLab/scicoqa",
     "url": "https://huggingface.co/spaces/UKPLab/scicoqa",
     "kind": "space",
     "note": "Space for the SciCoQA resource."
    }
   ],
   "papers": [
    {
     "title": "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
     "url": "https://arxiv.org/abs/1908.10084",
     "year": "2019"
    },
    {
     "title": "Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation",
     "url": "https://arxiv.org/abs/2004.09813",
     "year": "2020"
    },
    {
     "title": "GPL: Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval",
     "url": "https://arxiv.org/abs/2112.07577",
     "year": "2021"
    },
    {
     "title": "BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models",
     "url": "https://arxiv.org/abs/2104.08663",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Sentence-BERT",
     "url": "https://github.com/UKPLab/sentence-transformers",
     "note": "Sentence embedding library originated at UKP"
    },
    {
     "name": "GritHopper-7B",
     "url": "https://huggingface.co/UKPLab/GritHopper-7B",
     "note": "Multi-hop retrieval model"
    },
    {
     "name": "PeerQA",
     "url": "https://huggingface.co/datasets/UKPLab/PeerQA",
     "note": "Dataset for question answering on peer reviews"
    }
   ],
   "openness": "Code and datasets are public on GitHub and Hugging Face under mostly Apache-2.0.",
   "key_people": [
    {
     "name": "Iryna Gurevych",
     "role": "Founder and head, Professor at TU Darmstadt",
     "url": "https://en.wikipedia.org/wiki/Iryna_Gurevych",
     "note": "Founded the lab; Wikipedia attributes its founding to her."
    }
   ],
   "discord_note": null,
   "contact_note": "The lab's group website at TU Darmstadt is the place for general contact; the GitHub organisation takes issues per project.",
   "sources": [
    "https://www.ukp.tu-darmstadt.de",
    "https://huggingface.co/UKPLab",
    "https://github.com/UKPLab",
    "https://arxiv.org/abs/1908.10084"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "ukp-lab",
   "mlat": 49.89445,
   "mlon": 8.59301,
   "logo": true
  },
  {
   "orig_id": "ultralytics",
   "name": "Ultralytics",
   "alias": null,
   "full_name": "Ultralytics",
   "type": "startup",
   "founded": null,
   "country": "United States",
   "city": "Frederick",
   "lat": 39.4143,
   "lon": -77.4105,
   "website": "https://www.ultralytics.com",
   "huggingface": "https://huggingface.co/Ultralytics",
   "github": "https://github.com/ultralytics",
   "x": "https://x.com/ultralytics",
   "discord": "https://discord.com/invite/ultralytics",
   "linkedin": null,
   "contact_email": "hello@ultralytics.com",
   "summary": "Computer-vision company behind the YOLO model family and the ultralytics library, released under AGPL-3.0 with enterprise licensing.",
   "tagline": "Ultralytics develops the YOLO model family and ultralytics library, released under AGPL-3.0 with enterprise licensing.",
   "about": [
    "Ultralytics is a computer-vision company led by founder and CEO Glenn Jocher. Its ultralytics Python package provides YOLO models for detection, segmentation and other tasks.",
    "Models are distributed under AGPL-3.0 by default, with a paid Enterprise license for private deployment. YOLO11 was released in September 2024."
   ],
   "mission": "Ultralytics aims to make vision AI accessible through a widely used open library.",
   "notice": null,
   "known_for": [
    "YOLO model family",
    "ultralytics Python library",
    "AGPL-3.0 with enterprise license"
   ],
   "focus": [
    "object detection",
    "computer vision",
    "YOLO"
   ],
   "contributions": [
    {
     "title": "YOLOv5",
     "note": "Released by Glenn Jocher on 26 June.",
     "year": "2020"
    },
    {
     "title": "YOLOv8",
     "note": "Introduced; project moves to AGPL-3.0.",
     "year": "2023"
    },
    {
     "title": "YOLO11",
     "note": "Released in September.",
     "year": "2024"
    },
    {
     "title": "YOLO26",
     "note": "Next-generation release, per the company's blog.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "ultralytics",
     "url": "https://github.com/ultralytics/ultralytics",
     "note": "YOLO models library"
    },
    {
     "name": "yolov5",
     "url": "https://github.com/ultralytics/yolov5",
     "note": "YOLOv5 in PyTorch"
    },
    {
     "name": "yolov3",
     "url": "https://github.com/ultralytics/yolov3",
     "note": "YOLOv3 in PyTorch"
    }
   ],
   "hf_highlights": [
    {
     "name": "YOLOv8",
     "url": "https://huggingface.co/Ultralytics/YOLOv8",
     "kind": "model",
     "note": "YOLOv8"
    },
    {
     "name": "YOLO11",
     "url": "https://huggingface.co/Ultralytics/YOLO11",
     "kind": "model",
     "note": "YOLO11"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "ultralytics",
     "url": "https://github.com/ultralytics/ultralytics",
     "note": "YOLO models library"
    },
    {
     "name": "yolov5",
     "url": "https://github.com/ultralytics/yolov5",
     "note": "YOLOv5 in PyTorch"
    },
    {
     "name": "yolov3",
     "url": "https://github.com/ultralytics/yolov3",
     "note": "YOLOv3 in PyTorch"
    }
   ],
   "openness": "AGPL-3.0 code and weights with commercial licence available.",
   "key_people": [
    {
     "name": "Glenn Jocher",
     "role": "Founder and CEO",
     "url": "https://github.com/glenn-jocher",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the website.",
   "contact_note": "Email printed on website.",
   "sources": [
    "https://www.ultralytics.com/about",
    "https://github.com/ultralytics",
    "https://huggingface.co/Ultralytics"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "ultralytics",
   "logo": true
  },
  {
   "orig_id": "unsloth",
   "name": "Unsloth",
   "alias": null,
   "full_name": "Unsloth",
   "type": "startup",
   "founded": 2023,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://unsloth.ai",
   "huggingface": "https://huggingface.co/unsloth",
   "github": "https://github.com/unslothai",
   "x": null,
   "discord": "https://discord.gg/unsloth",
   "linkedin": null,
   "contact_email": null,
   "summary": "Open-source project and company making LLM fine-tuning and RL faster and lighter; also publishes quantised model uploads.",
   "tagline": "Unsloth is an open-source project and company making LLM fine-tuning and reinforcement learning faster and lighter.",
   "about": [
    "Unsloth was started by brothers Daniel and Michael Han, who launched it on 30 November 2023 as an open-source tool for faster single-GPU fine-tuning. They began in Sydney and later moved to San Francisco.",
    "It joined Y Combinator's Summer 2024 batch and also publishes quantized model uploads."
   ],
   "mission": "Unsloth aims to make training and running LLMs cheaper for individual developers.",
   "notice": null,
   "known_for": [
    "Fast LLM fine-tuning",
    "Dynamic quantization",
    "Open training kernels"
   ],
   "focus": [
    "fine-tuning",
    "training infrastructure",
    "quantization"
   ],
   "contributions": [
    {
     "title": "Unsloth launch",
     "note": "Introduced on 30 November as faster AI training.",
     "year": "2023"
    },
    {
     "title": "Y Combinator S24",
     "note": "The company joins YC.",
     "year": "2024"
    },
    {
     "title": "Seed funding",
     "note": "Seed round of roughly $500K, reported in September or October.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "unsloth",
     "url": "https://github.com/unslothai/unsloth",
     "note": "Flagship fine-tuning library; very high star count."
    },
    {
     "name": "notebooks",
     "url": "https://github.com/unslothai/notebooks",
     "note": "250+ fine-tuning and RL notebooks."
    },
    {
     "name": "hyperlearn",
     "url": "https://github.com/unslothai/hyperlearn",
     "note": "Faster, leaner ML algorithms."
    },
    {
     "name": "unsloth-zoo",
     "url": "https://github.com/unslothai/unsloth-zoo",
     "note": "Utilities for Unsloth."
    }
   ],
   "hf_highlights": [
    {
     "name": "Qwen3.8-27B-GGUF",
     "url": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF",
     "kind": "model",
     "note": "Top downloaded GGUF quantisation."
    },
    {
     "name": "OpenMathReasoning-mini",
     "url": "https://huggingface.co/datasets/unsloth/OpenMathReasoning-mini",
     "kind": "dataset",
     "note": "Small math reasoning dataset."
    },
    {
     "name": "LaTeX_OCR",
     "url": "https://huggingface.co/datasets/unsloth/LaTeX_OCR",
     "kind": "dataset",
     "note": "LaTeX OCR dataset for vision fine-tuning."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "unsloth",
     "url": "https://github.com/unslothai/unsloth",
     "note": "Fine-tuning library"
    },
    {
     "name": "notebooks",
     "url": "https://github.com/unslothai/notebooks",
     "note": "Training notebooks"
    },
    {
     "name": "unsloth-zoo",
     "url": "https://github.com/unslothai/unsloth-zoo",
     "note": "Utilities"
    }
   ],
   "openness": "Core library open source (Apache-2.0), plus GGUF/quantised releases on Hugging Face.",
   "key_people": [
    {
     "name": "Daniel Han",
     "role": "Co-founder",
     "url": "https://www.ai.engineer/speakers/daniel-han",
     "note": null
    },
    {
     "name": "Michael Han",
     "role": "Co-founder",
     "url": "https://github.com/unslothai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://unsloth.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://unsloth.ai",
    "https://huggingface.co/unsloth",
    "https://github.com/unslothai"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "unsloth",
   "mlat": 37.66599,
   "mlon": -122.48315,
   "logo": true
  },
  {
   "orig_id": "upstage",
   "name": "Upstage",
   "alias": null,
   "full_name": "Upstage",
   "type": "startup",
   "founded": 2020,
   "country": "South Korea",
   "city": "Seoul",
   "lat": 37.57,
   "lon": 126.98,
   "website": "https://upstage.ai",
   "huggingface": "https://huggingface.co/upstage",
   "github": "https://github.com/upstageai",
   "x": "https://twitter.com/upstageai",
   "discord": "https://discord.com/invite/upstage",
   "linkedin": null,
   "contact_email": "contact@upstage.ai",
   "summary": "Korean AI company building the Solar language models and document processing engines; some models released openly.",
   "tagline": "Upstage is a Korean AI company behind the Solar language models and document processing engines, some released openly.",
   "about": [
    "Upstage was founded in October 2020 by Sung Kim, a former professor at the Hong Kong University of Science and Technology who previously led Naver's Clova team. It builds language models and document AI for enterprises.",
    "Its compact Solar 10.7B model was released in December 2023, followed by Solar Pro 2 in July 2025. Some models are published openly, though later releases are under the Upstage Solar License."
   ],
   "mission": "Upstage aims to build efficient models and document AI for enterprises.",
   "notice": null,
   "known_for": [
    "Solar 10.7B",
    "Compact LLMs",
    "Document AI"
   ],
   "focus": [
    "language models",
    "documents",
    "Korean"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Founded by Sung Kim, formerly of HKUST and Naver's Clova team.",
     "year": "2020"
    },
    {
     "title": "Series A",
     "note": "31.6 billion won raised.",
     "year": "2021"
    },
    {
     "title": "Solar 10.7B",
     "note": "Compact model released.",
     "year": "2023"
    },
    {
     "title": "Series B",
     "note": "100 billion won raised from investors including SoftBank Ventures Asia and Korea Development Bank.",
     "year": "2024"
    },
    {
     "title": "Solar Pro 2",
     "note": "Released in July.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "dataverse",
     "url": "https://github.com/UpstageAI/dataverse",
     "note": "Data engineering framework for LLM data, the org's top repo."
    },
    {
     "name": "solar-prompt-cookbook",
     "url": "https://github.com/UpstageAI/solar-prompt-cookbook",
     "note": "Prompting examples for Solar models."
    },
    {
     "name": "evalverse",
     "url": "https://github.com/UpstageAI/evalverse",
     "note": "Unified framework for evaluating LLMs."
    },
    {
     "name": "cookbook",
     "url": "https://github.com/UpstageAI/cookbook",
     "note": "API examples and guides."
    }
   ],
   "hf_highlights": [
    {
     "name": "upstage/Solar-Open-100B",
     "url": "https://huggingface.co/upstage/Solar-Open-100B",
     "kind": "model",
     "note": "Most downloaded Upstage model."
    },
    {
     "name": "upstage/SOLAR-10.7B-Instruct-v1.0",
     "url": "https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0",
     "kind": "model",
     "note": "Instruction-tuned SOLAR 10.7B."
    },
    {
     "name": "upstage/SOLAR-10.7B-v1.0",
     "url": "https://huggingface.co/upstage/SOLAR-10.7B-v1.0",
     "kind": "model",
     "note": "Base depth-up-scaled model."
    },
    {
     "name": "upstage/dp-bench",
     "url": "https://huggingface.co/datasets/upstage/dp-bench",
     "kind": "dataset",
     "note": "Document parsing benchmark."
    }
   ],
   "papers": [
    {
     "title": "SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling",
     "url": "https://arxiv.org/abs/2312.15166",
     "year": "2023"
    }
   ],
   "notable_releases": [
    {
     "name": "Solar-Open2-250B",
     "url": "https://huggingface.co/upstage/Solar-Open2-250B",
     "note": "Model or dataset on its Hugging Face page"
    },
    {
     "name": "TinySolar-187m-4k",
     "url": "https://huggingface.co/upstage/TinySolar-187m-4k",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Selected open weights on Hugging Face; commercial API for others.",
   "key_people": [
    {
     "name": "Sung Kim",
     "role": "Co-founder and CEO",
     "url": "https://upstage.ai",
     "note": null
    }
   ],
   "discord_note": "Community server for Solar and Upstage API users, linked from upstage.ai.",
   "contact_note": "Upstage lists a contact address and a community Discord on its website.",
   "sources": [
    "https://huggingface.co/upstage",
    "https://upstage.ai"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "upstage",
   "mlat": 37.59165,
   "mlon": 126.93269,
   "logo": true
  },
  {
   "orig_id": "vector-institute",
   "name": "Vector Institute",
   "alias": null,
   "full_name": "Vector Institute",
   "type": "institute",
   "founded": 2017,
   "country": "Canada",
   "city": "Toronto",
   "lat": 43.6532,
   "lon": -79.3832,
   "website": "https://vectorinstitute.ai",
   "huggingface": null,
   "github": "https://github.com/VectorInstitute",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "info@vectorinstitute.ai",
   "summary": "Toronto nonprofit AI research institute publishing open tools for LLM inference, federated learning and health AI.",
   "tagline": "Vector Institute is a Toronto nonprofit AI research institute that also publishes open tools for LLM work, federated learning and health AI.",
   "about": [
    "Vector opened in March 2017 at MaRS in Toronto as an independent nonprofit. Geoffrey Hinton became chief scientific adviser, Richard Zemel research director, and Garth Gibson CEO.",
    "Its research is led by Hinton with research directors Daniel Roy and Graham Taylor as of 2023. Its GitHub organization hosts open engineering tools."
   ],
   "mission": "Vector aims to drive excellence in machine learning and deep learning in Canada.",
   "notice": null,
   "known_for": [
    "Deep learning research hub",
    "Hinton as scientific adviser",
    "Open AI engineering tools"
   ],
   "focus": [
    "tooling",
    "federated learning",
    "evaluation"
   ],
   "contributions": [
    {
     "title": "Institute opens",
     "note": "Launched at MaRS in Toronto.",
     "year": "2017"
    }
   ],
   "github_highlights": [
    {
     "name": "fed-rag",
     "url": "https://github.com/VectorInstitute/fed-rag",
     "note": "Fine-tuning RAG systems framework."
    },
    {
     "name": "vector-inference",
     "url": "https://github.com/VectorInstitute/vector-inference",
     "note": "Efficient LLM inference on Slurm clusters."
    },
    {
     "name": "cyclops",
     "url": "https://github.com/VectorInstitute/cyclops",
     "note": "Evaluate and monitor clinical AI models."
    },
    {
     "name": "vbll",
     "url": "https://github.com/VectorInstitute/vbll",
     "note": "Variational Bayesian last layers for uncertainty."
    }
   ],
   "hf_highlights": [],
   "papers": [],
   "notable_releases": [
    {
     "name": "vector-inference",
     "url": "https://github.com/VectorInstitute/vector-inference",
     "note": "Slurm LLM inference"
    },
    {
     "name": "fed-rag",
     "url": "https://github.com/VectorInstitute/fed-rag",
     "note": "Federated RAG"
    }
   ],
   "openness": "Open-source code on GitHub.",
   "key_people": [
    {
     "name": "Geoffrey Hinton",
     "role": "Chief Scientific Adviser",
     "url": "https://www.cs.toronto.edu/~hinton/",
     "note": null
    },
    {
     "name": "Richard Zemel",
     "role": "Research Director",
     "url": "https://www.cs.toronto.edu/~zemel/",
     "note": null
    },
    {
     "name": "Daniel Roy",
     "role": "Research Director",
     "url": "https://vectorinstitute.ai",
     "note": null
    },
    {
     "name": "Graham Taylor",
     "role": "Research Director",
     "url": "https://vectorinstitute.ai",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact details and links are on the lab's website (https://vectorinstitute.ai); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://vectorinstitute.ai",
    "https://github.com/VectorInstitute"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "vector-institute",
   "logo": true
  },
  {
   "orig_id": "vistec-airesearch",
   "name": "VISTEC-depa AI Research Institute of Thailand",
   "alias": null,
   "full_name": "VISTEC-depa AI Research Institute of Thailand",
   "type": "academic",
   "founded": null,
   "country": "Thailand",
   "city": "Rayong",
   "lat": 12.72,
   "lon": 101.15,
   "website": "https://airesearch.in.th/",
   "huggingface": "https://huggingface.co/airesearch",
   "github": "https://github.com/vistec-AI",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Thai institute with 200+ models and Thai datasets on Hugging Face.",
   "tagline": "VISTEC-depa AIRI is a Thai institute publishing Thai NLP datasets and models such as WangchanBERTa and WangchanLION.",
   "about": [
    "The institute was formed through a collaboration between depa and VISTEC. It provides large datasets and pretrained models for Thai.",
    "WangchanBERTa is its Thai transformer language model. WangchanLION, built with AI Singapore on SEA-LION, is released with code, data and weights under Apache 2.0, and WangchanLION-v3 continues pretraining from Llama 3.1 8B."
   ],
   "mission": "The institute aims to provide Thai language AI resources.",
   "notice": null,
   "known_for": [
    "WangchanBERTa",
    "WangchanLION",
    "Thai datasets"
   ],
   "focus": [
    "Thai",
    "datasets",
    "language models"
   ],
   "contributions": [
    {
     "title": "English-Thai translation dataset and models",
     "note": "Released in June.",
     "year": "2020"
    },
    {
     "title": "WangchanBERTa",
     "note": "Thai RoBERTa-based model pretrained on 78 GB of text, released in March.",
     "year": "2021"
    },
    {
     "title": "Thai Speech Emotion Dataset",
     "note": "Released in March.",
     "year": "2021"
    },
    {
     "title": "WangchanLION",
     "note": "Instruction-tuned Thai SEA-LION model.",
     "year": "2024"
    },
    {
     "title": "WangchanLION-v3",
     "note": "Continual pretraining from Llama 3.1 8B.",
     "year": "2025"
    }
   ],
   "github_highlights": [
    {
     "name": "thai2transformers",
     "url": "https://github.com/vistec-AI/thai2transformers",
     "note": "Pretraining Thai transformers."
    },
    {
     "name": "wav2vec2-large-xlsr-53-th",
     "url": "https://github.com/vistec-AI/wav2vec2-large-xlsr-53-th",
     "note": "Thai wav2vec2 fine-tuning."
    },
    {
     "name": "commonvoice-th",
     "url": "https://github.com/vistec-AI/commonvoice-th",
     "note": "Kaldi recipe for Thai Common Voice."
    },
    {
     "name": "dataset-releases",
     "url": "https://github.com/vistec-AI/dataset-releases",
     "note": "Dataset releases."
    }
   ],
   "hf_highlights": [
    {
     "name": "airesearch/wav2vec2-large-xlsr-53-th",
     "url": "https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th",
     "kind": "model",
     "note": "Thai ASR."
    },
    {
     "name": "airesearch/wangchanberta-base-att-spm-uncased",
     "url": "https://huggingface.co/airesearch/wangchanberta-base-att-spm-uncased",
     "kind": "model",
     "note": "WangchanBERTa."
    },
    {
     "name": "airesearch/WangchanThaiInstruct",
     "url": "https://huggingface.co/datasets/airesearch/WangchanThaiInstruct",
     "kind": "dataset",
     "note": "Thai instruction data."
    },
    {
     "name": "airesearch/thai-ser",
     "url": "https://huggingface.co/datasets/airesearch/thai-ser",
     "kind": "dataset",
     "note": "Thai speech emotion data."
    }
   ],
   "papers": [
    {
     "title": "WangchanBERTa: Pretraining transformer-based Thai Language Models",
     "url": "https://arxiv.org/abs/2101.09635",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Qwen3-8B-ReTool-w-CoBe",
     "url": "https://huggingface.co/airesearch/Qwen3-8B-ReTool-w-CoBe",
     "note": "Model or dataset on its Hugging Face page"
    }
   ],
   "openness": "Models and datasets on Hugging Face.",
   "key_people": [
    {
     "name": "Sarana Nutanong",
     "role": "Researcher at VISTEC, WangchanBERTa co-author",
     "url": "https://arxiv.org/abs/2101.09635",
     "note": "Author list of the WangchanBERTa paper."
    }
   ],
   "discord_note": null,
   "contact_note": "airesearch.in.th lists the institute's contacts.",
   "sources": [
    "https://huggingface.co/airesearch",
    "https://airesearch.in.th/"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "vistec-depa-ai-research-institute-of-thailand",
   "logo": true
  },
  {
   "orig_id": "ox-vgg",
   "name": "Visual Geometry Group",
   "alias": "Oxford",
   "full_name": "Visual Geometry Group (Oxford)",
   "type": "academic",
   "founded": null,
   "country": "United Kingdom",
   "city": "Oxford",
   "lat": 51.752,
   "lon": -1.2577,
   "website": "https://www.robots.ox.ac.uk/~vgg/",
   "huggingface": null,
   "github": "https://github.com/ox-vgg",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Computer vision research group in the Engineering Science department at the University of Oxford, long known for open datasets and tools such as VGGFace2, the VGG Image Annotator and the VGGT 3D reconstruction model.",
   "tagline": "Oxford's Visual Geometry Group, source of VGG networks and open vision datasets and tools.",
   "about": [
    "The Visual Geometry Group is within the Department of Engineering Science at Oxford, founded by Andrew Zisserman.",
    "The 2014 VGG paper by Simonyan and Zisserman placed first in localization and second in classification at ILSVRC 2014.",
    "GitHub hosts tools."
   ],
   "mission": "The group advances computer vision and shares data and software.",
   "notice": null,
   "known_for": [
    "VGG networks",
    "VGGFace2",
    "VGG Image Annotator",
    "VGGT"
   ],
   "focus": [
    "computer vision",
    "3D vision",
    "datasets"
   ],
   "contributions": [
    {
     "title": "VGG at ILSVRC",
     "note": "Very deep ConvNets placed first in localization.",
     "year": "2014"
    }
   ],
   "github_highlights": [
    {
     "name": "vgg_face2",
     "url": "https://github.com/ox-vgg/vgg_face2",
     "note": "Open-source repository. (712 stars)"
    },
    {
     "name": "via",
     "url": "https://github.com/ox-vgg/via",
     "note": "(MIRROR) see https://gitlab.com/vgg/via/ (249 stars)"
    },
    {
     "name": "vise",
     "url": "https://github.com/ox-vgg/vise",
     "note": "(MIRROR of https://gitlab.com/vgg/vise) VGG Image Search Engine (VISE) is a standalone software for visual sea (52 stars)"
    },
    {
     "name": "vgg_face_search",
     "url": "https://github.com/ox-vgg/vgg_face_search",
     "note": "(MIRROR) Face finding engine that runs on a local service. Includes a pipeline for preprocessing a user-define (21 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "VGGFace2: A dataset for recognising faces across pose and age",
     "url": "https://arxiv.org/abs/1710.08092",
     "year": "2017"
    },
    {
     "title": "VGGT: Visual Geometry Grounded Transformer",
     "url": "https://arxiv.org/abs/2503.11651",
     "year": "2025"
    }
   ],
   "notable_releases": [
    {
     "name": "VGGFace2",
     "url": "https://github.com/ox-vgg/vgg_face2",
     "note": "Large-scale face recognition dataset released with the VGGFace2 paper"
    },
    {
     "name": "VGG Image Annotator (VIA)",
     "url": "https://github.com/ox-vgg/via",
     "note": "Browser-based image, audio and video annotation tool (GitHub mirror)"
    },
    {
     "name": "VISE",
     "url": "https://github.com/ox-vgg/vise",
     "note": "Standalone image search engine for large collections"
    },
    {
     "name": "VGG Face search",
     "url": "https://github.com/ox-vgg/vgg_face_search",
     "note": "Face finding engine with a preprocessing pipeline"
    }
   ],
   "openness": "Datasets, annotation tools and research code are published openly on GitHub (mirrors of the group GitLab) and the group site, with permissive or research licences per project.",
   "key_people": [
    {
     "name": "Andrew Zisserman",
     "role": "Professor, founder",
     "url": "http://www.robots.ox.ac.uk/~az",
     "note": "Co-author of the VGG networks."
    },
    {
     "name": "Andrea Vedaldi",
     "role": "Professor",
     "url": "http://www.vlfeat.org/~vedaldi/",
     "note": "Principal researcher in the group."
    },
    {
     "name": "João Henriques",
     "role": "Principal researcher",
     "url": "http://www.robots.ox.ac.uk/~joao",
     "note": "Group researcher."
    },
    {
     "name": "Christian Rupprecht",
     "role": "Principal researcher",
     "url": "https://chrirupp.github.io",
     "note": "Group researcher."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact through the group website.",
   "sources": [
    "https://www.robots.ox.ac.uk/~vgg/",
    "https://github.com/ox-vgg",
    "https://arxiv.org/abs/1710.08092"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "visual-geometry-group",
   "mlat": 51.78736,
   "mlon": -1.2577,
   "logo": true
  },
  {
   "orig_id": "vita-epfl",
   "name": "VITA",
   "alias": "EPFL",
   "full_name": "VITA (EPFL)",
   "type": "academic",
   "founded": null,
   "country": "Switzerland",
   "city": "Lausanne",
   "lat": 46.5191,
   "lon": 6.5668,
   "website": "https://www.epfl.ch/labs/vita/",
   "huggingface": null,
   "github": "https://github.com/vita-epfl",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Visual Intelligence for Transportation lab at EPFL, publishing open code for pose estimation, trajectory prediction and video generation.",
   "tagline": "EPFL's lab on socially aware AI for transportation, with open pose and trajectory code.",
   "about": [
    "VITA, the Visual Intelligence for Transportation lab, is led by Alexandre Alahi at EPFL, where he joined in 2017. It works on socially aware AI for safe autonomous systems and mobility.",
    "Its open code includes OpenPifPaf for human pose estimation, which was combined with commercial licensing for technology transfer, along with trajectory prediction and video generation projects on GitHub."
   ],
   "mission": "The lab makes AI socially aware for transportation.",
   "notice": null,
   "known_for": [
    "Socially aware AI",
    "Human motion prediction"
   ],
   "focus": [
    "vision",
    "autonomous driving"
   ],
   "contributions": [
    {
     "title": "Alahi joins EPFL",
     "note": "Start of lab leadership.",
     "year": "2017"
    },
    {
     "title": "OpenPifPaf technology transfer",
     "note": "Open-source pose estimation tool paired with commercial licenses.",
     "year": "2022"
    },
    {
     "title": "Editor's choice award",
     "note": "Image and Vision Computing, for human motion prediction.",
     "year": "2021"
    },
    {
     "title": "TEDx talk on socially aware AI",
     "note": "Public outreach on the lab's research.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "Stable-Video-Infinity",
     "url": "https://github.com/vita-epfl/Stable-Video-Infinity",
     "note": "[ICLR 26 Oral] Stable Video Infinity: Infinite-Length Video Generation with Error Recycling (2,622 stars)"
    },
    {
     "name": "CrowdNav",
     "url": "https://github.com/vita-epfl/CrowdNav",
     "note": "[ICRA] Crowd-aware Robot Navigation (742 stars)"
    },
    {
     "name": "UniTraj",
     "url": "https://github.com/vita-epfl/UniTraj",
     "note": "A Unified Framework for scalable Vehicle Trajectory Prediction, ECCV 2024 (493 stars)"
    },
    {
     "name": "monoloco",
     "url": "https://github.com/vita-epfl/monoloco",
     "note": "A 3D vision library from 2D keypoints: monocular and stereo 3D detection for humans, social distancing, and bo (462 stars)"
    }
   ],
   "hf_highlights": [],
   "papers": [
    {
     "title": "OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association",
     "url": "https://arxiv.org/abs/2103.02440",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "Stable-Video-Infinity",
     "url": "https://github.com/vita-epfl/Stable-Video-Infinity",
     "note": "Infinite-length video generation (ICLR 2026 oral)"
    },
    {
     "name": "CrowdNav",
     "url": "https://github.com/vita-epfl/CrowdNav",
     "note": "Crowd-aware robot navigation"
    },
    {
     "name": "UniTraj",
     "url": "https://github.com/vita-epfl/UniTraj",
     "note": "Framework for vehicle trajectory prediction"
    },
    {
     "name": "monoloco",
     "url": "https://github.com/vita-epfl/monoloco",
     "note": "3D vision from 2D keypoints"
    }
   ],
   "openness": "Research code is public on GitHub for most papers.",
   "key_people": [
    {
     "name": "Alexandre Alahi",
     "role": "Head of lab",
     "url": "https://people.epfl.ch/alexandre.alahi",
     "note": null
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via site.",
   "sources": [
    "https://www.epfl.ch/labs/vita/",
    "https://github.com/vita-epfl",
    "https://arxiv.org/abs/2103.02440"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "vita",
   "mlat": 46.55446,
   "mlon": 6.5668,
   "logo": true
  },
  {
   "orig_id": "voxel51",
   "name": "Voxel51",
   "alias": null,
   "full_name": "Voxel51",
   "type": "startup",
   "founded": null,
   "country": "United States",
   "city": "Ann Arbor",
   "lat": 42.2808,
   "lon": -83.743,
   "website": "https://voxel51.com",
   "huggingface": "https://huggingface.co/Voxel51",
   "github": "https://github.com/voxel51",
   "x": "https://x.com/voxel51",
   "discord": "https://discord.com/invite/fiftyone-community",
   "linkedin": null,
   "contact_email": null,
   "summary": "Visual AI company behind the open-source FiftyOne toolkit for curating datasets and evaluating computer-vision models, with many datasets on Hugging Face.",
   "tagline": "Voxel51 builds the open-source FiftyOne toolkit for curating datasets and evaluating computer vision models.",
   "about": [
    "Voxel51 was co-founded by University of Michigan professor Jason Corso and Brian Moore and is based in Ann Arbor. Sources differ on whether it began in 2016 or 2018.",
    "FiftyOne open source launched in August 2020 and reached version 1.0 in October 2024. The commercial platform builds on it."
   ],
   "mission": "Voxel51 aims to help teams build better visual AI through data-centric tooling.",
   "notice": null,
   "known_for": [
    "FiftyOne",
    "Dataset curation",
    "Model evaluation"
   ],
   "focus": [
    "computer vision",
    "datasets",
    "evaluation tooling"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Co-founded in Ann Arbor by Brian Moore and Jason Corso; sources give 2016 or 2018.",
     "year": "2016"
    },
    {
     "title": "FiftyOne launch",
     "note": "Open-source experimentation tool announced.",
     "year": "2020"
    },
    {
     "title": "Series B",
     "note": "$30M round led by Bessemer Venture Partners, announced May 2024.",
     "year": "2024"
    },
    {
     "title": "FiftyOne 1.0",
     "note": "Open-source version 1.0 released in October.",
     "year": "2024"
    }
   ],
   "github_highlights": [
    {
     "name": "fiftyone",
     "url": "https://github.com/voxel51/fiftyone",
     "note": "Refine datasets and visual AI models"
    },
    {
     "name": "voxelgpt",
     "url": "https://github.com/voxel51/voxelgpt",
     "note": "Assistant for querying visual datasets"
    },
    {
     "name": "fiftyone-examples",
     "url": "https://github.com/voxel51/fiftyone-examples",
     "note": "Examples of using FiftyOne"
    }
   ],
   "hf_highlights": [
    {
     "name": "LVIS",
     "url": "https://huggingface.co/datasets/Voxel51/LVIS",
     "kind": "dataset",
     "note": "LVIS in FiftyOne format"
    },
    {
     "name": "Food101",
     "url": "https://huggingface.co/datasets/Voxel51/Food101",
     "kind": "dataset",
     "note": "Food101 in FiftyOne format"
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "fiftyone",
     "url": "https://github.com/voxel51/fiftyone",
     "note": "Refine datasets and visual AI models"
    },
    {
     "name": "voxelgpt",
     "url": "https://github.com/voxel51/voxelgpt",
     "note": "Assistant for querying visual datasets"
    },
    {
     "name": "fiftyone-examples",
     "url": "https://github.com/voxel51/fiftyone-examples",
     "note": "Examples of using FiftyOne"
    }
   ],
   "openness": "Apache-2.0 FiftyOne; large set of open datasets in the Hugging Face Hub.",
   "key_people": [
    {
     "name": "Jason Corso",
     "role": "Co-founder",
     "url": "https://voxel51.com/about",
     "note": null
    },
    {
     "name": "Brian Moore",
     "role": "Co-founder and CEO",
     "url": "https://voxel51.com/about",
     "note": null
    }
   ],
   "discord_note": "Community Discord linked from the website.",
   "contact_note": "Via website.",
   "sources": [
    "https://voxel51.com",
    "https://github.com/voxel51",
    "https://huggingface.co/Voxel51"
   ],
   "confidence": "medium",
   "region": "North America",
   "id": "voxel51",
   "logo": true
  },
  {
   "orig_id": "ytu-cosmos",
   "name": "YTU COSMOS AI Research Group",
   "alias": null,
   "full_name": "YTU COSMOS AI Research Group",
   "type": "academic",
   "founded": null,
   "country": "Turkey",
   "city": "Istanbul",
   "lat": 41.0082,
   "lon": 28.9784,
   "website": "https://huggingface.co/ytu-ce-cosmos",
   "huggingface": "https://huggingface.co/ytu-ce-cosmos",
   "github": null,
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "COSMOS AI research group at the Computer Engineering Department of Yildiz Technical University in Istanbul, publishing open Turkish language models, corpora and benchmarks.",
   "tagline": "Yildiz Technical University's COSMOS group, publishing open Turkish language models.",
   "about": [
    "COSMOS is in the Computer Engineering Department of Yildiz Technical University.",
    "It released cosmosGPT in 2024 and Turkish-Llama and Turkish-Gemma models.",
    "Models are on Hugging Face."
   ],
   "mission": "The group builds Turkish-first language models.",
   "notice": null,
   "known_for": [
    "cosmosGPT Turkish models",
    "Turkish-Llama",
    "Turkish-Gemma"
   ],
   "focus": [
    "Turkish LLMs"
   ],
   "contributions": [
    {
     "title": "cosmosGPT",
     "note": "Monolingual Turkish training paper.",
     "year": "2024"
    },
    {
     "title": "Turkish-Gemma-9b",
     "note": "9B model released on Hugging Face.",
     "year": "2025"
    },
    {
     "title": "Cosmos Turkish Corpus v1.0",
     "note": "Turkish pretraining corpus published as a dataset.",
     "year": "2025"
    },
    {
     "title": "ModernBERT-TR",
     "note": "Family of roughly 150M-parameter Turkish encoder models.",
     "year": "2026"
    }
   ],
   "github_highlights": [],
   "hf_highlights": [
    {
     "name": "ytu-ce-cosmos/Turkish-Gemma-9b-T1",
     "url": "https://huggingface.co/ytu-ce-cosmos/Turkish-Gemma-9b-T1",
     "kind": "model",
     "note": "Model with 181 likes on the Hub."
    },
    {
     "name": "ytu-ce-cosmos/Turkish-Llama-8b-v0.1",
     "url": "https://huggingface.co/ytu-ce-cosmos/Turkish-Llama-8b-v0.1",
     "kind": "model",
     "note": "Model with 60 likes on the Hub."
    },
    {
     "name": "ytu-ce-cosmos/Cosmos-Turkish-Corpus-v1.0",
     "url": "https://huggingface.co/datasets/ytu-ce-cosmos/Cosmos-Turkish-Corpus-v1.0",
     "kind": "dataset",
     "note": "Dataset with 28 likes on the Hub."
    }
   ],
   "papers": [],
   "notable_releases": [
    {
     "name": "Turkish-Gemma-9b-T1",
     "url": "https://huggingface.co/ytu-ce-cosmos/Turkish-Gemma-9b-T1",
     "note": "Turkish reasoning model built on Gemma"
    },
    {
     "name": "Turkish-Llama-8b",
     "url": "https://huggingface.co/ytu-ce-cosmos/Turkish-Llama-8b-v0.1",
     "note": "Turkish-adapted Llama model"
    },
    {
     "name": "Cosmos Turkish Corpus",
     "url": "https://huggingface.co/datasets/ytu-ce-cosmos/Cosmos-Turkish-Corpus-v1.0",
     "note": "Turkish text corpus"
    },
    {
     "name": "gsm8k_tr",
     "url": "https://huggingface.co/datasets/ytu-ce-cosmos/gsm8k_tr",
     "note": "Turkish translation of GSM8K"
    }
   ],
   "openness": "Model weights and datasets are public on Hugging Face; some checkpoints inherit Gemma or Llama licence terms.",
   "key_people": [
    {
     "name": "Mehmet Fatih Amasyali",
     "role": "Professor, group member listed on Hugging Face",
     "url": "https://huggingface.co/ytu-ce-cosmos",
     "note": "Listed as a member of the organization."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via university.",
   "sources": [
    "https://huggingface.co/ytu-ce-cosmos",
    "https://huggingface.co/ytu-ce-cosmos"
   ],
   "confidence": "medium",
   "region": "Europe",
   "id": "ytu-cosmos-ai-research-group",
   "logo": true
  },
  {
   "orig_id": "zai-zhipu",
   "name": "Z.ai",
   "alias": "Zhipu AI",
   "full_name": "Z.ai (Zhipu AI)",
   "type": "startup",
   "founded": 2019,
   "country": "China",
   "city": "Beijing",
   "lat": 39.9,
   "lon": 116.41,
   "website": "https://z.ai",
   "huggingface": "https://huggingface.co/zai-org",
   "github": "https://github.com/zai-org",
   "x": "https://twitter.com/Zai_org",
   "discord": null,
   "linkedin": null,
   "contact_email": null,
   "summary": "Builder of the ChatGLM and GLM model families, also working on agents, evaluation and graph research. Releases weights and training or inference techniques.",
   "tagline": "Z.ai, formerly Zhipu AI, is a Tsinghua spinoff that releases the GLM models, including MIT-licensed open weights.",
   "about": [
    "Zhipu AI is a Tsinghua University spinoff that rebranded as Z.ai in 2025. Co-founder Tang Jie is chief scientist, and the company listed in Hong Kong on 8 January 2026.",
    "It released GLM-5, a 744B open-weight model under MIT, in February 2026, and later GLM-5.2."
   ],
   "mission": "Z.ai aims to build frontier general models and release open weights.",
   "notice": null,
   "known_for": [
    "GLM model family",
    "GLM-5 open weights",
    "ChatGLM"
   ],
   "focus": [
    "language models",
    "agents",
    "evaluation",
    "Chinese",
    "English"
   ],
   "contributions": [
    {
     "title": "Company founded",
     "note": "Tsinghua spin-out founded in Beijing.",
     "year": "2019"
    },
    {
     "title": "GLM paper",
     "note": "First GLM training algorithm presented at ACL.",
     "year": "2022"
    },
    {
     "title": "MIT-licensed releases and rebrand",
     "note": "Models released under MIT in 2025; GLM-4.5 and international rebrand to Z.ai in July.",
     "year": "2025"
    },
    {
     "title": "Hong Kong listing",
     "note": "Listed on 8 January.",
     "year": "2026"
    },
    {
     "title": "GLM-5",
     "note": "Released 12 February; GLM-5.1 open-sourced in April.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "ChatGLM-6B",
     "url": "https://github.com/zai-org/ChatGLM-6B",
     "note": "Open bilingual dialogue model, the org's most starred repo."
    },
    {
     "name": "Open-AutoGLM",
     "url": "https://github.com/zai-org/Open-AutoGLM",
     "note": "Open phone agent model and framework."
    },
    {
     "name": "ChatGLM3",
     "url": "https://github.com/zai-org/ChatGLM3",
     "note": "Third generation of ChatGLM."
    },
    {
     "name": "CogVideo",
     "url": "https://github.com/zai-org/CogVideo",
     "note": "Text and image to video generation (CogVideoX)."
    }
   ],
   "hf_highlights": [
    {
     "name": "GLM-5.3-Flash",
     "url": "https://huggingface.co/zai-org/GLM-5.3-Flash",
     "kind": "model",
     "note": "Most downloaded model in the org."
    },
    {
     "name": "GLM-OCR",
     "url": "https://huggingface.co/zai-org/GLM-OCR",
     "kind": "model",
     "note": "OCR model."
    },
    {
     "name": "GLM-5.3",
     "url": "https://huggingface.co/zai-org/GLM-5.3",
     "kind": "model",
     "note": "Flagship GLM release."
    },
    {
     "name": "LongBench-v2",
     "url": "https://huggingface.co/datasets/zai-org/LongBench-v2",
     "kind": "dataset",
     "note": "Long-context benchmark."
    }
   ],
   "papers": [
    {
     "title": "ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools",
     "url": "https://arxiv.org/abs/2406.12793",
     "year": "2024"
    },
    {
     "title": "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models",
     "url": "https://arxiv.org/abs/2508.06471",
     "year": "2025"
    },
    {
     "title": "CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer",
     "url": "https://arxiv.org/abs/2408.06072",
     "year": "2024"
    },
    {
     "title": "GLM-130B: An Open Bilingual Pre-trained Model",
     "url": "https://arxiv.org/abs/2210.02414",
     "year": "2022"
    }
   ],
   "notable_releases": [
    {
     "name": "GLM-5.3",
     "url": "https://huggingface.co/zai-org/GLM-5.3",
     "note": "GLM-5.3"
    },
    {
     "name": "GLM-5.3-Flash",
     "url": "https://huggingface.co/zai-org/GLM-5.3-Flash",
     "note": "GLM-5.3 Flash"
    }
   ],
   "openness": "Open weights on Hugging Face and open code on GitHub; licences per model.",
   "key_people": [
    {
     "name": "Tang Jie",
     "role": "Co-founder, chief scientist",
     "url": "https://en.wikipedia.org/wiki/Zhipu_AI",
     "note": "Tsinghua professor; named as co-founder on the company's Wikipedia entry."
    }
   ],
   "discord_note": null,
   "contact_note": "Contact via the official site, GitHub issues or Hugging Face discussions.",
   "sources": [
    "https://huggingface.co/zai-org",
    "https://z.ai"
   ],
   "confidence": "high",
   "region": "Asia",
   "id": "z-ai",
   "mlat": 39.88273,
   "mlon": 116.3407,
   "logo": true
  },
  {
   "orig_id": "zyphra",
   "name": "Zyphra",
   "alias": null,
   "full_name": "Zyphra",
   "type": "startup",
   "founded": 2020,
   "country": "United States",
   "city": "San Francisco",
   "lat": 37.7749,
   "lon": -122.4194,
   "website": "https://www.zyphra.com",
   "huggingface": "https://huggingface.co/Zyphra",
   "github": "https://github.com/Zyphra",
   "x": null,
   "discord": "https://discord.com/invite/puMPr32s2t",
   "linkedin": null,
   "contact_email": null,
   "summary": "San Francisco AI lab releasing open-weight models including Zonos text-to-speech and the Zamba hybrid language models, plus an EEG foundation model.",
   "tagline": "Zyphra is an AI lab in Palo Alto releasing open-weight models, including the ZAYA1-8B reasoning mixture-of-experts model.",
   "about": [
    "Zyphra says it was founded in 2020 and lists members in Palo Alto, London and Montreal. Dealroom names Krithik Puthalath and Danny Martinelli as founders.",
    "It has released ZAYA1-8B, a reasoning-focused mixture-of-experts model trained on AMD Instinct MI300 GPUs, and demonstrated large-scale training on a full AMD stack."
   ],
   "mission": "Zyphra aims at full-stack, efficient AI models.",
   "notice": null,
   "known_for": [
    "ZAYA1-8B",
    "AMD-trained models",
    "Open-weight releases"
   ],
   "focus": [
    "speech",
    "language models",
    "state-space models",
    "neuroscience"
   ],
   "contributions": [
    {
     "title": "Zonos-v0.1",
     "note": "Open-weight 1.6B text-to-speech model under Apache 2.0.",
     "year": "2025"
    },
    {
     "title": "ZAYA1 on AMD",
     "note": "Large-scale mixture-of-experts training on AMD Instinct MI300X with IBM Cloud.",
     "year": "2025"
    },
    {
     "title": "ZAYA1-8B",
     "note": "Open reasoning model released under Apache 2.0.",
     "year": "2026"
    }
   ],
   "github_highlights": [
    {
     "name": "Zonos",
     "url": "https://github.com/Zyphra/Zonos",
     "note": "Zonos-v0.1 open-weight TTS model."
    },
    {
     "name": "zuna",
     "url": "https://github.com/Zyphra/zuna",
     "note": "Python package for the Zuna EEG foundation model."
    },
    {
     "name": "ZONOS2",
     "url": "https://github.com/Zyphra/ZONOS2",
     "note": "Open-weight TTS mixture-of-experts."
    },
    {
     "name": "BlackMamba",
     "url": "https://github.com/Zyphra/BlackMamba",
     "note": "Code for BlackMamba."
    }
   ],
   "hf_highlights": [
    {
     "name": "ZAYA1-VL-8B",
     "url": "https://huggingface.co/Zyphra/ZAYA1-VL-8B",
     "kind": "model",
     "note": "Most downloaded Zyphra model."
    },
    {
     "name": "ZONOS3-Mini",
     "url": "https://huggingface.co/Zyphra/ZONOS3-Mini",
     "kind": "model",
     "note": "Compact voice-cloning TTS."
    },
    {
     "name": "Zyda-2",
     "url": "https://huggingface.co/datasets/Zyphra/Zyda-2",
     "kind": "dataset",
     "note": "Large open pretraining dataset."
    },
    {
     "name": "Zyda",
     "url": "https://huggingface.co/datasets/Zyphra/Zyda",
     "kind": "dataset",
     "note": "First Zyda dataset."
    }
   ],
   "papers": [
    {
     "title": "Zamba: A Compact 7B SSM Hybrid Model",
     "url": "https://arxiv.org/abs/2405.16712",
     "year": "2024"
    },
    {
     "title": "BlackMamba: Mixture of Experts for State-Space Models",
     "url": "https://arxiv.org/abs/2402.01771",
     "year": "2024"
    }
   ],
   "notable_releases": [
    {
     "name": "Zonos",
     "url": "https://github.com/Zyphra/Zonos",
     "note": "Open-weight TTS"
    },
    {
     "name": "BlackMamba",
     "url": "https://github.com/Zyphra/BlackMamba",
     "note": "SSM-MoE code"
    },
    {
     "name": "ZONOS3-Mini",
     "url": "https://huggingface.co/Zyphra/ZONOS3-Mini",
     "note": "Compact TTS"
    }
   ],
   "openness": "Open weights, many under Apache-2.0.",
   "key_people": [
    {
     "name": "Krithik Puthalath",
     "role": "Founder",
     "url": "https://www.zyphra.com",
     "note": null
    },
    {
     "name": "Danny Martinelli",
     "role": "Founder",
     "url": "https://www.zyphra.com",
     "note": null
    }
   ],
   "discord_note": "Public Discord linked from the lab's own website for community discussion and announcements.",
   "contact_note": "Contact details and links are on the lab's website (https://www.zyphra.com); code questions go through issues on their GitHub organisation.",
   "sources": [
    "https://www.zyphra.com",
    "https://huggingface.co/Zyphra",
    "https://github.com/Zyphra"
   ],
   "confidence": "high",
   "region": "North America",
   "id": "zyphra",
   "mlat": 37.65771,
   "mlon": -122.45204,
   "logo": true
  },
  {
   "orig_id": "ufal",
   "name": "ÚFAL",
   "alias": "Charles University",
   "full_name": "ÚFAL (Charles University)",
   "type": "academic",
   "founded": 1990,
   "country": "Czechia",
   "city": "Prague",
   "lat": 50.0755,
   "lon": 14.4378,
   "website": "https://ufal.mff.cuni.cz",
   "huggingface": "https://huggingface.co/ufal",
   "github": "https://github.com/ufal",
   "x": null,
   "discord": null,
   "linkedin": null,
   "contact_email": "ufal@ufal.mff.cuni.cz",
   "summary": "Institute of Formal and Applied Linguistics at Charles University, a long-standing NLP lab behind UDPipe, MorphoDiTa and Czech language models.",
   "tagline": "ÚFAL is Charles University's Institute of Formal and Applied Linguistics, behind UDPipe and Czech treebanks.",
   "about": [
    "ÚFAL was established in 1990 at the Faculty of Mathematics and Physics, continuing the Laboratory of Algebraic Linguistics active since the early 1960s.",
    "It develops UDPipe, a trainable pipeline for tokenization, tagging, lemmatization and parsing under MPL 2.0, and the Prague Dependency Treebank family."
   ],
   "mission": "The institute advances computational linguistics and language resources.",
   "notice": null,
   "known_for": [
    "UDPipe",
    "Prague Dependency Treebank",
    "Czech NLP"
   ],
   "focus": [
    "NLP",
    "Czech",
    "parsing",
    "machine translation"
   ],
   "contributions": [
    {
     "title": "Institute established",
     "note": "ÚFAL founded.",
     "year": "1990"
    },
    {
     "title": "UDPipe",
     "note": "Pipeline released.",
     "year": "2016"
    },
    {
     "title": "whisper_streaming",
     "note": "Real-time streaming wrapper for Whisper transcription.",
     "year": "2023"
    }
   ],
   "github_highlights": [
    {
     "name": "whisper_streaming",
     "url": "https://github.com/ufal/whisper_streaming",
     "note": "Whisper realtime streaming for long speech-to-text transcription and translation (3,679 stars)"
    },
    {
     "name": "SimulStreaming",
     "url": "https://github.com/ufal/SimulStreaming",
     "note": "Simultaneous streaming speech system (697 stars)"
    },
    {
     "name": "udpipe",
     "url": "https://github.com/ufal/udpipe",
     "note": "Trainable pipeline for tokenising, tagging, lemmatising and parsing (426 stars)"
    },
    {
     "name": "neuralmonkey",
     "url": "https://github.com/ufal/neuralmonkey",
     "note": "Sequence learning tool for NLP on TensorFlow (417 stars)"
    }
   ],
   "hf_highlights": [
    {
     "name": "ufal/vit-historical-page",
     "url": "https://huggingface.co/ufal/vit-historical-page",
     "kind": "model",
     "note": "Vision transformer for historical page classification."
    },
    {
     "name": "ufal/robeczech-base",
     "url": "https://huggingface.co/ufal/robeczech-base",
     "kind": "model",
     "note": "Czech RoBERTa model."
    },
    {
     "name": "ufal/parczech4speech-segmented",
     "url": "https://huggingface.co/datasets/ufal/parczech4speech-segmented",
     "kind": "dataset",
     "note": "Segmented Czech parliamentary speech data."
    },
    {
     "name": "ufal/eleczech-lc-small",
     "url": "https://huggingface.co/ufal/eleczech-lc-small",
     "kind": "model",
     "note": "Small Czech ELECTRA model."
    }
   ],
   "papers": [
    {
     "title": "Turning Whisper into Real-Time Transcription System",
     "url": "https://arxiv.org/abs/2307.14743",
     "year": "2023"
    },
    {
     "title": "RobeCzech: Czech RoBERTa, a monolingual contextualized language representation model",
     "url": "https://arxiv.org/abs/2105.11314",
     "year": "2021"
    }
   ],
   "notable_releases": [
    {
     "name": "UDPipe",
     "url": "https://github.com/ufal/udpipe",
     "note": "Trainable pipeline for Universal Dependencies"
    },
    {
     "name": "MorphoDiTa",
     "url": "https://github.com/ufal/morphodita",
     "note": "Morphological dictionary and tagger"
    },
    {
     "name": "RobeCzech",
     "url": "https://huggingface.co/ufal/robeczech-base",
     "note": "Czech RoBERTa model"
    }
   ],
   "openness": "Tools and models are public on GitHub and Hugging Face, mostly under MPL or MIT-style licences.",
   "key_people": [
    {
     "name": "Jan Hajič",
     "role": "Professor, Editor-in-Chief of the Prague Bulletin of Mathematical Linguistics",
     "url": "https://ufal.mff.cuni.cz/~hajic",
     "note": "Leads Charles University's group in Humane-AI-Net."
    },
    {
     "name": "Ondřej Bojar",
     "role": "Researcher in machine translation",
     "url": "https://ufal.mff.cuni.cz/~bojar/publications",
     "note": "Machine translation research."
    }
   ],
   "discord_note": null,
   "contact_note": "The institute's home page prints a general contact address and postal details.",
   "sources": [
    "https://ufal.mff.cuni.cz",
    "https://huggingface.co/ufal",
    "https://github.com/ufal"
   ],
   "confidence": "high",
   "region": "Europe",
   "id": "fal",
   "mlat": 50.0322,
   "mlon": 14.4378,
   "logo": true
  }
 ]
}