
Community collective · Beijing, China
OpenBMB / ModelBest
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.
Beijing, China
- Founded
- 2021
- Type
- Community collective
- Based in
- Beijing, China
- Key papers
- 3
- Featured projects
- 8
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 and ambitionsOpenBMB aims to build foundation models and systems toward AGI, with efficient models that run on consumer devices.
Known for
- 01
MiniCPM on-device language models
- 02
MiniCPM-V multimodal models
- 03
VoxCPM text-to-speech
- 04
UltraRAG framework
Milestones
- 2022
ModelBest founded
The company was spun out of Tsinghua's NLP Laboratory in August 2022.
- 2023
Luca assistant
ModelBest launched its multimodal dialogue assistant Luca in November 2023.
- 2025
MiniCPM 4.0
Released in June 2025 alongside the CPM.cu inference framework.
- 2026
MiniCPM-o 4.5
A full-duplex multimodal model for video, audio and text input introduced in early 2026.
- 2026
MiniCPM5-2B
A two-billion-parameter model whose local deployment tests were published in November 2026.
- 2024
MiniCPM paper
The MiniCPM small language models were described in an April 2024 arXiv paper with 25 authors.
Open source
Tokenizer-free multilingual TTS.
Multi-agent software development.
On-device multimodal models.
On-device LLMs.
- MiniCPM5-2Bmodel
Small on-device LLM.
- VoxCPM2model
Multilingual TTS.
MoE multimodal model.
- Ultra-FineWebdataset
Filtered web corpus.
Research papers
Notable releases
Openness
Open weights on Hugging Face and code on GitHub.
People
- Liu ZhiyuanProfessor, Tsinghua University; ModelBest co-founder and Chief Scientist
Associated with the Tsinghua NLP group and ModelBest; co-author of the MiniCPM paper.
- Maosong SunProfessor, Tsinghua University
Senior author of the MiniCPM paper; described as a principal investigator of the community.
- Li DahaiCEO, ModelBest
Co-author of the MiniCPM paper and head of the company.
- Shengding HuFirst author of MiniCPM
Led the MiniCPM small language model paper.
Listed from the lab's own pages and public profiles.
Contact and community
Contact via its website or GitHub and Hugging Face pages.
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.



