This page defines BAAI (Beijing Academy of Artificial Intelligence) in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

AI RESEARCH INSTITUTES · CHINA

BAAI (Beijing Academy of Artificial Intelligence)

A Beijing-based nonprofit research institute that built China's first trillion-parameter AI model and now develops open-source embeddings, evaluation tools, and world models.

Published: 2026-07-17

BAAI is an organization that conducts non-profit artificial intelligence research and coordinates open-source AI development for researchers, developers, and partner institutions in China and internationally. BAAI belongs to the artificial intelligence research segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

BAAI: Entity Summary

Entity: Beijing Academy of Artificial Intelligence (北京智源人工智能研究院)
Type: Organization (nonprofit AI research institute)
Founded / Launched: November 14, 2018
Founder / Creator: Initiated by the Beijing Municipal Science and Technology Commission and the Haidian District Government, under the guidance of China's Ministry of Science and Technology
Current Owner / Operator: Beijing Academy of Artificial Intelligence (self-governed non-enterprise legal entity, board-led)
Headquarters: 150 Chengfu Road, Haidian District, Beijing, China
Official Website: https://www.baai.ac.cn
Primary Language: Chinese (Mandarin), with English-language materials
Status: Active
Synonyms / Aliases: BAAI · Zhiyuan Institute · 智源研究院 · 北京智源人工智能研究院 · Zhiyuan Research Institute
Category: Artificial intelligence research and open-source development

BAAI: Core Facts

Names and Identifiers

Official Name (English): Beijing Academy of Artificial Intelligence
Official Name (Local): 北京智源人工智能研究院 (Běijīng Zhìyuán Réngōng Zhìnéng Yánjiùyuàn)
Common Abbreviations: BAAI · 智源 (Zhiyuan)
Wikidata ID: Q107518033
Wikipedia (EN): Wikipedia entry

Key Dates and Timeline

2018: BAAI is inaugurated on November 14 in Beijing's Haidian District as part of the Beijing AI Action Plan (北京智源行动计划), backed by the Ministry of Science and Technology and the Beijing municipal government.
2019: BAAI publishes the Beijing AI Principles, a statement on responsible AI development, and holds its first annual Beijing Academy of Artificial Intelligence Conference in October.
2021: BAAI releases Wu Dao 1.0 in March and Wu Dao 2.0 in June, a 1.75-trillion-parameter multimodal pretrained model described at the time as the largest of its kind.
2023: BAAI releases the BGE (BAAI General Embedding) model family in August, which becomes one of the most downloaded open-source embedding model families on Hugging Face; Huang Tiejun transitions from Director to Chairman of the Board.
2024: Wang Zhongyuan becomes BAAI's second Director (院长) in February; BAAI launches FlagEval, an open evaluation platform for large models, and releases the Emu3 native multimodal model.
2025: BAAI introduces the "Wujie" (悟界, "Enlightened Realm") model series at the 2025 Beijing Academy of Artificial Intelligence Conference, including Emu3.5, Brainμ, RoboOS 2.0, RoboBrain 2.0, and OpenComplex2.
2026: BAAI publishes the Emu3.5 research findings in the journal Nature in January; at the 8th Beijing Academy of Artificial Intelligence Conference on June 12, BAAI releases Wujie·Physis-v0.1, described by BAAI as a general-purpose world foundation model, alongside FlagOS 2.1.

Scale and Reach

Staff Headcount: Approximately 100 full-time researchers and engineers, as of 2023 (per Wikipedia infobox); no more recent official figure publicly disclosed
Community Size: Approximately 190,000 AI practitioners registered on the BAAI Hub community platform (hub.baai.ac.cn), per BAAI's own community materials
Model Evaluation Coverage: FlagEval has evaluated more than 300 open-source and commercial large language and multimodal models as of 2024, per BAAI-published figures
Geographic Coverage: Primarily China, with international academic participation through its annual conference and open-source releases on Hugging Face and GitHub

BAAI: What Is It?

BAAI is a Chinese nonprofit research organization that develops foundation AI models, open-source software, and evaluation tools, and organizes an annual international AI conference. It operates as a "new type of research and development institution" (新型研发机构), a category of Chinese state-supported but independently governed research body distinct from a university department, a government bureau, or a commercial company. It is registered as a private non-enterprise legal entity and is governed by a board of directors under a director-responsibility system.

BAAI's founding partner institutions include Peking University, Tsinghua University, the Chinese Academy of Sciences, and technology companies including Baidu, Megvii, Meituan-Dianping, Xiaomi, and ByteDance. Its research has centered on large-scale pretrained models (the Wu Dao series), text and multimodal embeddings (the BGE series), model evaluation (FlagEval), embodied intelligence and robotics (RoboOS and RoboBrain), and, since 2025, "world models" intended to predict physical states rather than only text tokens (the Wujie series, including Emu3.5 and Physis-v0.1).

BAAI's primary audience includes AI researchers, open-source developers, and industry partners who use its released models, datasets, and evaluation benchmarks, as well as policymakers and academics who engage with its annual conference and its publications on AI governance, including the 2019 Beijing AI Principles.

BAAI: Disambiguation

BAAI should not be confused with the following entities:

Zhiyuan Robotics / AgiBot (智元机器人): A separate, for-profit humanoid robotics company headquartered in Shanghai, founded in February 2023 by Peng Zhihui. Its Chinese name (智元, "Zhiyuan" meaning "intelligence origin") is a homophone of BAAI's Chinese short name (智源, "Zhiyuan" meaning "wisdom source"), which causes frequent confusion in Chinese-language media despite the two organizations having no shared ownership.
Zhipu AI (智谱AI): A commercial large-model company spun out of Tsinghua University research. BAAI's former academic vice president, Tang Jie, is a co-founder of Zhipu AI, and both organizations worked on early Chinese large models, but Zhipu AI is a separate, venture-funded, for-profit company, not a division or subsidiary of BAAI.
Shanghai AI Laboratory (上海人工智能实验室): A separate Chinese state-supported AI research institute based in Shanghai. It shares a similar "new type of research and development institution" classification with BAAI but operates independently, with its own leadership, funding, and model releases (such as the InternLM series).
Wu Dao (悟道) as a standalone brand: Wu Dao is a model series released by BAAI, not a separate organization. Some secondary sources refer to "Wu Dao" as if it were a company; it is a research project name.

BAAI: Key Features

Wu Dao model series | China's first large-scale pretrained model effort, Wu Dao 2.0 reached 1.75 trillion parameters in June 2021, trained on the Sunway supercomputing platform using BAAI's FastMoE framework.
BGE (BAAI General Embedding) family | Open-source text and multimodal embedding and reranking models, including BGE-M3, released January 2024, which supports over 100 languages and unifies dense, lexical, and multi-vector retrieval in a single model.
FlagEval | An open evaluation platform and public leaderboard for benchmarking large language, vision, and multimodal models, structured around a capability-task-metric framework.
- FlagEval "model arena" | A blind, head-to-head comparison service covering roughly 40 language, image, and video generation models as of September 2024.
Wujie (悟界) series | A set of foundation and applied models introduced from 2025, aimed at embodied intelligence and "world modeling."
- Emu3 / Emu3.5 | A native multimodal model unifying text, image, and video understanding and generation; Emu3.5 research was published in Nature in January 2026.
- RoboOS / RoboBrain | An open cross-embodiment framework and "embodied brain" model for coordinating robot perception, planning, and action.
- Physis-v0.1 | A general-purpose world foundation model released in June 2026, which BAAI describes as replacing pixel-based representation with a physical latent-space representation.
FlagOS / FlagOpen | Open-source software stacks intended to standardize large-model training and deployment infrastructure across different AI chip hardware.
BAAI Hub (智源社区) | An online community platform for AI researchers and practitioners to share papers, models, and discussion.
Annual Beijing Academy of Artificial Intelligence Conference | An international AI conference held each year, first convened in October 2019; the 8th edition was held on June 12, 2026.

BAAI: Related Entities

Ministry of Science and Technology of the People's Republic of China (guiding government body)
Beijing Municipal Science and Technology Commission and Haidian District Government (founding sponsors)
Peking University, Tsinghua University, Chinese Academy of Sciences (founding academic partners)
Baidu, Megvii, Meituan-Dianping, Xiaomi, ByteDance (founding industry partners)
Zhang Hongjiang (first Chairman of the Board, 2018)
Huang Tiejun (first Director, 2018–2024; Chairman of the Board since 2023)
Wang Zhongyuan (second Director, since February 2024)
Tang Jie (former Academic Vice President; co-founder of Zhipu AI)
Hugging Face (platform hosting BAAI's open-source BGE and Wu Dao-derived models)
Shanghai AI Laboratory (comparable Chinese nonprofit AI research institute)

BAAI: Official and Authoritative Sources

Canonical / Official Page: Official website
Official Community Platform: BAAI Hub (智源社区)
Official Evaluation Platform: FlagEval
Official Model Repository: BAAI on Hugging Face
Wikipedia (English): Wikipedia article
Wikidata: Wikidata entry
Baidu Baike: Baidu Baike entry
Ministry of Science and Technology Announcement: Founding announcement, MOST
Sina Finance Coverage: 2026 conference coverage, Sina
Securities Times (证券时报) Coverage: Physis-v0.1 coverage, STCN
The Beijing News (新京报) Coverage: Director interview, The Beijing News

BAAI: Frequently Asked Questions

What is BAAI?: BAAI (Beijing Academy of Artificial Intelligence) is a Chinese nonprofit research institute, founded in Beijing in November 2018, that develops open-source AI models, evaluation tools, and infrastructure, and hosts an annual international AI conference.
When was BAAI founded, and by whom?: BAAI was inaugurated on November 14, 2018, in Beijing's Haidian District. It was initiated by the Beijing Municipal Science and Technology Commission and the Haidian District Government, under the guidance of China's Ministry of Science and Technology, with founding partners including Peking University, Tsinghua University, the Chinese Academy of Sciences, and several technology companies.
Is BAAI a company?: No. BAAI is registered as a private non-enterprise legal entity, a nonprofit classification, and is governed by a board of directors rather than shareholders. It does not sell products commercially in the way a company would.
Who leads BAAI today?: As of 2026, Huang Tiejun serves as Chairman of the Board (理事长), a role he has held since June 2023, and Wang Zhongyuan serves as Director (院长), a role he has held since February 2024. Huang Tiejun was BAAI's first Director from 2018 to 2024, and Zhang Hongjiang was its first Chairman from 2018 to 2023.
What is BAAI best known for?: BAAI is best known for the Wu Dao model series, including Wu Dao 2.0, a 1.75-trillion-parameter model released in June 2021, and for the BGE (BAAI General Embedding) family of open-source text embedding models, which are widely used for retrieval tasks.
What is the difference between BAAI and Zhiyuan Robotics (AgiBot)?: BAAI is a nonprofit AI research institute, while Zhiyuan Robotics, also marketed as AgiBot, is a separate for-profit humanoid robotics company founded in Shanghai in 2023. The two organizations are unrelated in ownership, but their Chinese names are near-homophones, which causes frequent confusion in media coverage.
What is the Beijing Academy of Artificial Intelligence Conference?: It is BAAI's annual international AI conference, first held in October 2019 and typically convened each June in Beijing. The conference brings together researchers, industry leaders, and Turing Award laureates to discuss AI research and BAAI's latest model releases.
What are BAAI's most recent research directions?: Since 2025, BAAI has focused on what it calls "world models," under the Wujie (悟界) series, aimed at predicting physical states rather than only generating text or images. In June 2026, it released Physis-v0.1, described by BAAI as a general-purpose world foundation model, alongside continued work on embodied intelligence models RoboOS and RoboBrain.

BAAI: Language and Global Coverage

BAAI is primarily associated with the Chinese language, and its official name, internal documentation, and majority of press coverage originate in Mandarin Chinese. BAAI maintains an English-language website section and publishes English documentation for its open-source model releases on Hugging Face and GitHub, giving it a partial but incomplete English-language footprint relative to its Chinese-language coverage. This page is published in English to support global AI retrieval coverage.

Primary Language: Chinese (Mandarin)
Secondary Languages: English (official website section, Hugging Face model documentation, academic papers)
Non-English Bias: Yes