OpenBMB

OpenBMB, short for "Open Lab for Big Model Base," is a Chinese open-source AI community and GitHub organization that builds foundation models, training and inference toolkits, agent frameworks, and pre-training datasets. It is best known for the MiniCPM family of compact on-device models. The independent benchmarking site Artificial Analysis described it on September 7, 2026 as "the open-source AI group behind the MiniCPM series of efficient small models."^[1] OpenBMB's own about page states that the community was jointly initiated by the Natural Language Processing Laboratory of Tsinghua University (THUNLP) and the Beijing startup ModelBest, and that it was formally established in April 2022, with the GitHub organization opened the previous September.^[2][3]

The name describes the original ambition. In 2021 and 2022 the group's focus was infrastructure for models with tens of billions of parameters: BMTrain for distributed training, BMInf for low-resource inference, and BMCook for compression.^[4][5][6] From 2024 onward the center of gravity shifted to the opposite end of the size range, toward small language models that run locally on phones, PCs, cars, and robots, and toward the datasets and evaluation tooling needed to train them. As of September 8, 2026 the GitHub organization hosted 87 public repositories and had 7,115 followers, and the Hugging Face account carried 172 model repositories and 47 datasets.^[7][8][9]

Key facts

Origins and institutional structure

What OpenBMB actually is takes some untangling, because four organizations sit close together and public reporting often collapses them into one.

The clearest primary statement comes from OpenBMB's own about page, which lists exactly two founder logos, THUNLP and ModelBest, and dates the community's formal establishment to April 2022.^[2] An archived Chinese version of the same page from December 2023 says the same thing in plainer language: "The OpenBMB open-source community was jointly supported and initiated by the Tsinghua University Natural Language Processing Laboratory and ModelBest."^[3] Both READMEs of the two flagship repositories carry an "Institutions" section that names the developing bodies: MiniCPM-V credits THUNLP and ModelBest; MiniCPM credits ModelBest, THUNLP, and the Gaoling School of Artificial Intelligence at Renmin University of China.^[4][5] The 2024 MiniCPM paper makes the same split visible in its byline: its 25 authors carry exactly two affiliations, "Department of Computer Science and Technology, Tsinghua University" and "Modelbest Inc."^[33]

The practical picture that emerges is this. OpenBMB is not a legal entity in its own right as far as any primary source shows. It is a shared publishing and community brand: the GitHub organization, the Hugging Face account, the openbmb.cn website, and a set of Discord and Feishu channels through which a Tsinghua lab and a company release work jointly. The researchers are employed either by Tsinghua University or by ModelBest, and at least one person, Liu Zhiyuan, holds a post on both sides. The commercial entity is separate and clearly identified. Caixin gives ModelBest's registered name as Beijing Mianbi Intelligent Technology Co. Ltd., notes that ModelBest is the brand name, and says the company originated in a natural-language-processing lab at Tsinghua and was formally established in August 2022.^[10] The South China Morning Post also dates the founding to August 2022.^[11]

The Beijing Academy of Artificial Intelligence (BAAI) belongs to the prehistory rather than the governance. OpenBMB's milestone timeline records that the community's core team took part in BAAI's "WuDao" large-model project and released CPM-1, a 2-billion-parameter Chinese model, in 2020, and the archived 2023 about page says the CPM-1, CPM-2, and CPM-3 models were developed under the Beijing Academy of Artificial Intelligence, reaching up to 198 billion parameters.^[2][3] Neither the current about page nor the founder logos list BAAI as an initiator or governing body of OpenBMB itself. Public sources therefore support "the team came out of the WuDao effort" but do not support "BAAI governs OpenBMB," and this article does not assert the latter.

A fourth distinction matters for anyone reading the repository list. THUNLP maintains its own separate GitHub organization at github.com/thunlp, described as the "Natural Language Processing Lab at Tsinghua University," with 273 public repositories.^[12] Two projects that OpenBMB's website has historically listed among its toolkits, OpenPrompt and OpenDelta, live in the thunlp organization rather than the OpenBMB one; the archived about pages of December 2023 and November 2025 both give the toolkit list as BMTrain, BMCook, BMInf, OpenPrompt, OpenDelta, and ModelCenter, and github.com/OpenBMB/OpenPrompt and github.com/OpenBMB/OpenDelta both return 404.^[3][12] UltraChat is a different case: it is a dataset rather than a toolkit, its repository sits under thunlp, and the data itself is published under the openbmb account on Hugging Face.^[9][12] Project attribution on this cluster is genuinely split between the two orgs, and a claim that a given repository "is OpenBMB's" should be checked against the actual URL.

People

Liu Zhiyuan (刘知远), an associate professor in Tsinghua's Department of Computer Science and Technology and a member of THUNLP, is the most visible figure associated with OpenBMB, and is also a co-founder and chief scientist of ModelBest; ModelBest's own newsroom refers to him with that title.^[13] See Liu Zhiyuan for a fuller biography.

Sun Maosong (孙茂松), the senior Tsinghua professor who supervised Liu's doctorate, appears as the last author on the OpenBMB toolkit papers, including the ACL 2022 system demonstration for BMInf.^[14] Han Xu (韩旭), also of THUNLP, is the first author of that paper; the full author list is Xu Han, Guoyang Zeng, Weilin Zhao, Zhiyuan Liu, Zhengyan Zhang, Jie Zhou, Jun Zhang, Jia Chao, and Maosong Sun.^[14]

On the company side, ModelBest's published material names Li Dahai (李大海) as co-founder and chief executive, and Lei Shengtao (雷升涛) as chief operating officer.^[13] Zeng Guoyang (曾国洋), a co-author on the BMInf paper, is quoted by the South China Morning Post as speaking for ModelBest on compute constraints: "Domestic resources are limited, talent isn't an issue, and data can be resolved, but the bottleneck lies in compute power."^[11][14]

The big model base software stack

The original OpenBMB projects were three complementary systems libraries, all still maintained and all under Apache 2.0.^[4][6][15]

The website's toolkit page has also listed OpenPrompt (prompt learning) and OpenDelta (parameter-efficient or "delta" tuning) alongside these, though both repositories are maintained under the thunlp organization.^[3][12]

The MiniCPM model family

MiniCPM is the project the community is now known for. The first release, MiniCPM-2B, arrived on February 1, 2024 with roughly 2.4 billion non-embedding parameters, and its model card described it as jointly open-sourced by ModelBest and the Tsinghua NLP lab.^[17] The line then split into a text branch, a vision branch (MiniCPM-V), an omni-modal branch (MiniCPM-o), a speech branch (VoxCPM), and, from 2026, an embodied branch (MiniCPM-Robot).^[4][5][18]

Sources: MiniCPM and MiniCPM-V repository changelogs, OpenBMB's milestone timeline, and Hugging Face repository metadata.^[2][4][5][8] The sizes are the labels OpenBMB used in the release each row describes, and for several models OpenBMB's own documents disagree with one another; the Hugging Face figures quoted alongside them are safetensors tensor totals rather than parameter counts. MiniCPM reconciles the published figures model by model.

MiniCPM-V carried 26,317 GitHub stars on September 8, 2026, second among the model repositories to VoxCPM at 36,842, and its current head model, MiniCPM-V 4.6, is described in the repository as a 1.3-billion-parameter total model deployable on iOS, Android, and HarmonyOS.^[5][7] MiniCPM-o 4.5, released in February 2026, is a 9-billion-parameter end-to-end model whose repository claims it "approaches Gemini 2.5 Flash in vision, speech, and full-duplex multimodal live streaming," a self-reported comparison rather than an independent one.^[5] The newest text model, MiniCPM5-2B, was released on September 7, 2026.^[1][4]

The whole line is built for edge AI rather than cloud serving, and the community's own framing is that capability per parameter is the thing to optimize. The MiniCPM4 technical report is titled "MiniCPM4: Ultra-Efficient LLMs on End Devices."^[19]

Two 2024 side branches are worth noting because they are cited elsewhere. On April 11, 2024 the project released MiniCPM-2B-128k, MiniCPM-1B, and MiniCPM-MoE-8x2B, the last of which the MiniCPM paper says was "initialized utilizing Sparse Upcycling" from a dense MiniCPM checkpoint, giving 13.6 billion non-embedding parameters with two of eight experts activated per token.^[4][33] See sparse upcycling. The design and training methodology of the early family were written up in "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies," posted to arXiv in April 2024 by Shengding Hu and 24 co-authors.^[33]

Licensing history

The MiniCPM license has changed twice, so a single blanket label would be wrong. At launch in February 2024 the README released the repository under Apache 2.0 but bound the model weights to the General Model License (GML), a Chinese license family that OpenBMB itself publishes: weights were free for academic research, and commercial use required writing to ModelBest for a certificate of authorization.^[20][21] By September 2024 the weights had moved to a bespoke "MiniCPM Model License," free for research and free for commercial use after filling in a registration questionnaire.^[22] The restriction was dropped entirely alongside the MiniCPM4 release: commit 71245495, dated June 6, 2025 and titled "change license," rewrote the section to a single line, "This repository and MiniCPM models are released under the Apache-2.0 License," which is where it stands today.^[4][22] MiniCPM-V, MiniCPM-o, and VoxCPM currently carry Apache 2.0 on both weights and code, and the MiniCPM5 model cards on Hugging Face are tagged apache-2.0.^[5][8][18] Not every repository in the organization is Apache-licensed: PilotDeck and StaffDeck are AGPL-3.0, and several research repositories carry MIT or no declared license.^[7]

Agent projects

OpenBMB was an unusually early and prolific publisher of LLM agent frameworks. Its own timeline groups XAgent, AgentVerse, and ChatDev under June 2023 as covering single-agent, multi-agent, and agent-application territory respectively, though the three GitHub repositories were actually created between May and October of that year.^[2][7]

ChatDev is the most-starred of these and is frequently discussed alongside MetaGPT, a separate Chinese multi-agent software-company framework that is not an OpenBMB project. Star counts and repository creation dates in the table above come from the GitHub API on September 8, 2026.^[7]

Datasets and evaluation

The dataset releases are a contribution in their own right, and several of them are reused far outside OpenBMB's own models. UltraChat, released incrementally from April 2023, is a large synthetic multi-turn instruction corpus that became a widely used supervised fine-tuning set in the open weights ecosystem; its GitHub repository sits under thunlp while the dataset is published on the openbmb Hugging Face account.^[9][12] UltraFeedback followed in September 2023: about 64,000 prompts, four responses each from different models, and GPT-4 annotations along four axes (instruction-following, truthfulness, honesty, helpfulness), for roughly 256,000 rated samples, released together with the UltraRM reward model and UltraCM critique model.^[29] The pair became a common recipe for DPO-style preference training on small open models, in the same lineage of data-quality-first arguments as the LIMA paper.

Since February 2026 these have been organized under a public "UltraData" platform with an L0 to L4 tiered data-management framework, in which Ultra-FineWeb sits as the L2 selected layer for general web data.^[30]

On evaluation, two OpenBMB benchmarks are widely cited. InfiniteBench (published as "∞Bench") extends long-context evaluation beyond 100,000 tokens across 12 tasks mixing real and synthetic material.^[31] OlympiadBench, accepted to ACL 2024, is a bilingual multimodal benchmark of olympiad-level mathematics and physics problems.^[32] The organization also maintains UltraEval, an ACL 2024 demo framework for evaluating foundation models, and UltraEval-Audio for audio evaluation.^[7]

Reception and independent evaluation

Most of the performance numbers on the MiniCPM repositories are self-reported: OpenCompass averages the team ran itself, or head-to-head comparisons against a set of models the authors chose. Those should not be read as independent evaluation.

The clearest third-party read comes from Artificial Analysis, which evaluated MiniCPM5-2B on its Intelligence Index v4.2 on September 7, 2026. It reported a score of 15, calling it "the highest Intelligence Index of any open weights model under 4B total parameters" and four points clear of Granite 4.2 3B at 11, while noting weaker results on knowledge and coding: 9 percent on Humanity's Last Exam, 9 percent on Terminal-Bench v2.1, and 0 percent on CritPt. Artificial Analysis also observed that the model's AA-Omniscience score of -12 "is earned by abstaining from answering rather than accuracy," since it attempts only 29 percent of questions.^[1] The same article is a good illustration of why index versions have to be named: the model scored 23 on the earlier Intelligence Index v4.1.1, and Artificial Analysis states that "its new score of 15 on v4.2 reflects the updated evaluation mix and weightings, and scores across the two versions are not directly comparable."^[1]

Adoption is easier to measure than quality. The per-repository counts the Hugging Face API reported on September 8, 2026 for the openbmb account sum to roughly 4.5 million model downloads over the preceding 30 days, led by MiniCPM-o 4.5 (753,030), MiniCPM5-1B (724,212), MiniCPM-V 4.6 (530,314), VoxCPM2 (410,797), MiniCPM-V 4.5 (264,825), and MiniCPM-o 2.6 (259,050).^[8] On GitHub the largest repositories on the same date were VoxCPM (36,842 stars), ChatDev (34,233), MiniCPM-V (26,317), MiniCPM (10,536), and XAgent (8,540).^[7] The models are carried by mainstream local-inference tooling: MiniCPM-V 4.6 was merged into Ollama's official library in June 2026, and MiniCPM-V 4.5 gained official llama.cpp, vLLM, and LLaMA-Factory support in September 2025.^[5]

OpenBMB's own about page lists Intel, MediaTek, NVIDIA, and Qualcomm as chip partners, and shows member-unit logos labelled THUHCSI, Open knowledge graph, BIAA, Northeastern University, and Zhihu.^[2] Its milestone entry for late 2024 names Huawei Cloud, MediaTek, Intel, Baidu Smart Cloud, Great Wall Motors, and Accelerated Evolution Robots among integration partners.^[2] These are the community's own claims about its partners rather than independently confirmed commercial relationships.

Current status

OpenBMB remains actively developed. In the first nine months of 2026 alone the community shipped ChatDev 2.0, AgentCPM-Explore, AgentCPM-Report, and UltraRAG 3.0 in January; MiniCPM-o 4.5 and the UltraData platform in February; the EdgeClaw device-cloud agent framework in March; VoxCPM2 in April; MiniCPM-V 4.6, MiniCPM5-1B, and PilotDeck in May; and MiniCPM5-2B in September.^[2][4]

The commercial arm has moved toward a public listing. On August 12, 2026 Citic Securities filed with China's securities regulator that it had been hired to guide ModelBest's pre-IPO tutoring for a mainland listing; the company has not disclosed a target size, timeline, or venue, and the mandatory tutoring process takes at least three months.^[11] Caixin reported that the move could make ModelBest the first large language model developer to list under the Shanghai STAR Market's regime for unprofitable technology companies.^[10] What that means for the open-source community's governance has not been stated publicly by either party, and this article does not speculate.

OpenBMB has also started running open competitions around its own architectures. The MiniCPM-SALA model card advertises a 2026 "Sparse Operator Acceleration & Race" (SOAR) run with SGLang and NVIDIA, with a prize pool the card puts above 100,000 US dollars, challenging entrants to optimize 9B-scale, million-token inference on NVIDIA 6000D hardware.^[34]

Recent work also points outward from language models. The organization has published MiniCPM-Robot, a vision-language-action model series for robots; DeepThinkVLA and SimpleNav for embodied research; ForgeTrain, described on ModelBest's site as a training framework written autonomously by AI; and SciCore-Omics and SciCore-Mol in computational biology.^[7][13] Whether those lines mature into flagships or stay research repositories is not yet determinable from public information.

References

1. ^1 ^2 ^3 ^4Artificial Analysis. "OpenBMB releases MiniCPM5-2B." Published September 7, 2026. artificialanalysis.ai/...nbmb-releases-minicpm5-2b
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9OpenBMB. "About Us" page content (milestones, founder logos, chip partners), served via openbmb.cn/...about . Retrieved September 8, 2026. openbmb.cn/about-us
3. ^1 ^2 ^3 ^4 ^5 ^6OpenBMB. "关于我们" (About Us), archived December 3, 2023. web.archive.org/...about-us
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10OpenBMB. "MiniCPM" repository README. Retrieved September 8, 2026. github.com/...MiniCPM
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9OpenBMB. "MiniCPM-V" repository README. Retrieved September 8, 2026. github.com/...MiniCPM-V
6. ^1 ^2 ^3OpenBMB. "BMCook: Model Compression for Big Models." github.com/...BMCook
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13GitHub REST API. Organization and repository metadata for github.com/OpenBMB. Retrieved September 8, 2026. api.github.com/...OpenBMB
8. ^1 ^2 ^3 ^4 ^5Hugging Face API. Model list for author `openbmb`. Retrieved September 8, 2026. huggingface.co/...models
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7Hugging Face API. Dataset list for author `openbmb`. Retrieved September 8, 2026. huggingface.co/...datasets
10. ^1 ^2Gu Zhaowei. "ModelBest Begins Pre-IPO Tutoring as On-Device AI Competition Heats Up." Caixin Global, August 12, 2026. caixinglobal.com/...competition-heats-up-102473564
11. ^1 ^2 ^3Xinmei Shen. "Chinese AI start-up ModelBest, partner to Samsung and Huawei, launches pre-IPO process." South China Morning Post, August 12, 2026. scmp.com/...-ipo-tutoring-process-mainland-listing
12. ^1 ^2 ^3 ^4 ^5GitHub. THUNLP organization ("Natural Language Processing Lab at Tsinghua University"). Retrieved September 8, 2026. github.com/thunlp
13. ^1 ^2 ^3ModelBest. Company website and newsroom (English and Chinese editions). Retrieved September 8, 2026. modelbest.cn/en
14. ^1 ^2 ^3 ^4Xu Han, Guoyang Zeng, Weilin Zhao, Zhiyuan Liu, Zhengyan Zhang, Jie Zhou, Jun Zhang, Jia Chao, Maosong Sun. "BMInf: An Efficient Toolkit for Big Model Inference and Tuning." Proceedings of ACL 2022: System Demonstrations, pages 224-230. aclanthology.org/2022.acl-demo.22
15. ^1 ^2OpenBMB. "BMTrain: Efficient Training for Big Models." github.com/...BMTrain
16. ^OpenBMB. "ModelCenter." github.com/...ModelCenter
17. ^OpenBMB. Model card, `openbmb/MiniCPM-2B-sft-bf16`. huggingface.co/...MiniCPM-2B-sft-bf16
18. ^1 ^2OpenBMB. "VoxCPM" repository README. github.com/...VoxCPM
19. ^MiniCPM Team. "MiniCPM4: Ultra-Efficient LLMs on End Devices." arXiv:2506.07900. arxiv.org/...2506.07900
20. ^OpenBMB. "MiniCPM" README at commit b3358343, February 2024 (General Model License terms). github.com/...README-en.md
21. ^OpenBMB. "General-Model-License" repository. github.com/...General-Model-License
22. ^1 ^2OpenBMB. "MiniCPM" README-en.md commit history, in particular commits b3358343 (February 19, 2024), c58c9a28 (September 28, 2024) and 71245495 ("change license," June 6, 2025). Retrieved September 8, 2026. github.com/...README-en.md
23. ^OpenBMB. "ChatDev 2.0 - DevAll" repository README. github.com/...ChatDev
24. ^OpenBMB. "XAgent" repository README. github.com/...XAgent
25. ^Yujia Qin et al. "ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs." arXiv:2307.16789; repository at github.com/...ToolBench
26. ^Weize Chen et al. "AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors." arXiv:2308.10848. github.com/...AgentVerse
27. ^Weize Chen et al. "Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence." arXiv:2407.07061. github.com/...IoA
28. ^OpenBMB, THUNLP, NEUIR, AI9stars. "UltraRAG" repository README. github.com/...UltraRAG
29. ^1 ^2Ganqu Cui et al. "UltraFeedback: Boosting Language Models with Scaled AI Feedback." arXiv:2310.01377; repository README at github.com/...UltraFeedback
30. ^1 ^2 ^3 ^4OpenBMB. "Ultra-FineWeb" dataset card and technical report (arXiv:2505.05427). huggingface.co/...Ultra-FineWeb
31. ^Xinrong Zhang et al. "∞Bench: Extending Long Context Evaluation Beyond 100K Tokens." arXiv:2402.13718; repository at github.com/...InfiniteBench
32. ^Chaoqun He et al. "OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems." ACL 2024; arXiv:2402.14008. github.com/...OlympiadBench
33. ^1 ^2 ^3Shengding Hu et al. "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies." arXiv:2404.06395, April 2024. arxiv.org/...2404.06395
34. ^OpenBMB. Model card, `openbmb/MiniCPM-SALA` (SOAR competition notice). huggingface.co/...MiniCPM-SALA
35. ^OpenBMB. Website header and page metadata strings (tagline "Big Model for Everyone" / 让大模型飞入千家万户), served in openbmb.cn/...main-DMog5HPc.js . Retrieved September 8, 2026. openbmb.cn