
Community collective · Bucharest, Romania
OpenLLM-Ro
Community of Romanian researchers releasing open Romanian language models, datasets, benchmarks and training code.
Bucharest, Romania
- Founded
- 2024
- Type
- Community collective
- Based in
- Bucharest, Romania
- Key papers
- 3
- Featured projects
- 8
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 and ambitionsOpenLLM-Ro aims to build and share open language models and resources for Romanian.
Known for
- 01
RoLlama, RoMistral and RoGemma Romanian models
- 02
RoMTBench and RoCulturaBench
- 03
Open Romanian training and evaluation code
Milestones
- 2024
First Romanian models
RoLlama2 models were released on 14 May 2024.
- 2024
RoMistral, RoGemma and RoLlama3
Additional model families were released between May and June 2024.
- 2024
Training code and benchmarks
Training code (August 2024) and RoMTBench with RoCulturaBench (August 2024) were published.
- 2025
Pretraining data and quality classifier
Augmented Romanian pretraining data was released in November 2025.
- 2026
Vision-language resources
Code, training and evaluation data for Romanian vision-language models were released in June 2026.
Open source
Recipes used for training Romanian models (23 stars)
Fork with code and data used for training Romanian models (11 stars)
Fork for evaluating Romanian LLMs with LLM-as-a-judge (2 stars)
Fork with code and data for evaluating Romanian LLMs (1 stars)
Romanian instruct model based on Llama 3.1.
Romanian instruct model based on Gemma 2.
- OpenLLM-Ro/ro_mmludataset
Romanian MMLU evaluation set.
- OpenLLM-Ro/fineweb2-ro-bertdataset
Romanian pretraining data with quality annotations.
Research papers
Notable releases
Openness
Models, datasets and evaluation code are public on Hugging Face and GitHub.
People
- Mihai MasalaLead author, "Vorbeşti Româneşte?"
First author of the RoLlama recipe paper (EMNLP 2024 Findings).
- Traian RebedeaResearcher; co-author
Senior author of the Romanian pretraining data quality paper.
- Mihai DascaluResearcher; co-author
Co-author of the founding paper.
- Marius LeordeanuResearcher; co-author
Co-author of the founding paper.
Listed from the lab's own pages and public profiles.
Contact and community
The site invites people to join the community and links to its GitHub and Hugging Face pages; no email is printed.
Sources
Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.



