BAAI (Beijing Academy of AI)

BAAI (Beijing Academy of Artificial Intelligence) is a nonprofit artificial intelligence research institute headquartered in Beijing, China, founded in November 2018 with backing from the Beijing Municipal Government, the Chinese Ministry of Science and Technology, and a coalition of Chinese universities and technology companies including Tsinghua University, Peking University, the Chinese Academy of Sciences, Baidu, Alibaba, Tencent, ByteDance, Xiaomi, and Megvii. It develops the Wu Dao foundation model series (notably Wu Dao 2.0, a 1.75-trillion-parameter multimodal model released June 2021), the FlagOpen open-source ecosystem (including FlagAI, FlagPerf, FlagEval, and other open-source AI infrastructure), and ongoing research across foundation models, multimodal AI, embodied AI, AI for science, and AI safety. As of April 2026, BAAI is one of the principal Chinese AI research bodies and the host institution for the annual BAAI Conference, the leading academic AI conference in China.

At a glance

• Founded: November 2018 in Beijing, China.
• Status: Nonprofit research institute. Backing from the Beijing Municipal Government, the Chinese Ministry of Science and Technology, and a coalition of Chinese universities and technology companies.
• Funding: Chinese government funding through the Beijing Municipal Government and the Ministry of Science and Technology, plus contributions from member companies. Specific budget allocations have not been publicly disclosed.
• Director: Zhongyuan Wang, Director (Chairman) of BAAI. Chinese AI research leader.
• Other notable leadership: Hang Li, former chief scientist at Bytedance and prominent NLP researcher (associated with BAAI through 2023). Jie Tang, Chief Scientist of Wu Dao 2.0 and Tsinghua University professor. Bowen Zhou, former director of BAAI's Foundation Model Research Center; later founded Z.AI (Zhipu AI).
• Open weights: Yes, partial. Selected models from the Wu Dao series and the FlagOpen ecosystem are released open-source. The principal Wu Dao 2.0 weights are not publicly released.
• Flagship outputs: Wu Dao 1.0 (March 2021), Wu Dao 2.0 (June 2021, 1.75 trillion parameters), Aquila (Aquila-7B and Aquila-33B open-weights bilingual foundation models), the FlagOpen open-source ecosystem (FlagAI, FlagPerf, FlagEval, FlagAttention, FlagEmbedding), the BAAI Conference (annual academic AI conference in Beijing), and research output across foundation models, multimodal AI, embodied AI, AI for science, and AI safety.

Origins

BAAI was founded in November 2018 by the Beijing Municipal Government in coordination with the Chinese Ministry of Science and Technology, with a mandate to advance artificial intelligence research and serve as a coordinating body for the Chinese AI research community. The founding member organizations included Tsinghua University, Peking University, the Chinese Academy of Sciences, Beihang University, Renmin University of China, the Beijing Institute of Technology, Baidu, Alibaba, Tencent, ByteDance, Xiaomi, and Megvii.

The 2018 to 2020 founding period focused on building research capacity and convening Chinese AI research talent across member institutions. BAAI launched the BAAI Conference in 2019 as an annual academic conference, with participation from international AI researchers and Chinese AI research community members.

The 2021 release of the Wu Dao series was BAAI's most consequential international research output. Wu Dao 1.0 (March 2021) was a Chinese-and-English foundation model with parameter count. Wu Dao 2.0 (June 2021) was a 1.75-trillion-parameter mixture-of-experts multimodal model that briefly held the title of the largest published foundation model globally. The Wu Dao 2.0 release attracted considerable international attention as a public demonstration of Chinese AI research capability at frontier scale. Jie Tang, a Tsinghua University professor, served as chief scientist of the Wu Dao 2.0 program.

The 2022 to 2024 period saw BAAI broaden its research output. The Aquila series of open-weights bilingual foundation models (Aquila-7B and Aquila-33B) was released through 2023 and 2024. The FlagOpen ecosystem launched a series of open-source AI infrastructure projects: FlagAI (a foundation-model training and serving framework), FlagPerf (an AI hardware-benchmarking framework), FlagEval (a foundation-model evaluation framework), and other projects. Several BAAI-affiliated researchers spun out commercial AI ventures: Bowen Zhou, the former director of BAAI's Foundation Model Research Center, founded Z.AI (Zhipu AI) in 2019, with Wu Dao alumni continuing into 2024 and 2026.

The 2024 to 2026 period has continued BAAI's research output across embodied AI, AI for science, and AI safety, alongside the continued FlagOpen ecosystem expansion. The BAAI Conference has continued as the principal Chinese academic AI conference annually.

Mission and strategy

BAAI's stated mission is to advance artificial intelligence research and innovation, with emphasis on foundation models, multimodal AI, embodied AI, AI for science, and AI safety. The strategic premise reflects the Chinese government commitment to artificial intelligence as a strategic-priority research area, with BAAI explicitly positioned as a coordinating body for the Chinese AI research community across academic and industrial member institutions.

The strategy combines four threads. First, the foundation model research program through the Wu Dao series and other research lines. Second, the FlagOpen open-source ecosystem providing AI infrastructure (training frameworks, benchmarking frameworks, evaluation frameworks) for the broader Chinese AI research community. Third, the convening role through the annual BAAI Conference and other research-coordination activities. Fourth, multi-discipline research output across foundation models, multimodal AI, embodied AI, AI for science, and AI safety.

The competitive premise is that BAAI's coordinating role across Chinese government, academic, and industrial institutions, combined with the Beijing AI research talent base and the Chinese government research-funding commitment, provide a durable structural advantage for BAAI as one of the principal Chinese AI research bodies.

Distribution channels include open-source distribution of the FlagOpen ecosystem and selected Wu Dao and Aquila variants through GitHub and Hugging Face, the BAAI Conference annually, published research output across major academic venues, and continued cross-institution research-coordination relationships.

Models and products

• Wu Dao series. Wu Dao 1.0 (March 2021, multilingual foundation model), Wu Dao 2.0 (June 2021, 1.75-trillion-parameter mixture-of-experts multimodal model). The principal high-profile foundation model series internationally. Wu Dao 2.0 weights are not publicly released.
• Aquila series. Aquila-7B and Aquila-33B (open-weights bilingual foundation models with Chinese and English capability), released through 2023 and 2024. Open-weights through Hugging Face.
• FlagOpen ecosystem. FlagAI (foundation-model training and serving framework), FlagPerf (AI hardware-benchmarking framework), FlagEval (foundation-model evaluation framework), FlagAttention, FlagEmbedding, and other open-source AI infrastructure projects. Open-source through GitHub.
• BAAI Conference. Annual academic AI conference in Beijing.
• Multi-discipline research output. Published research output across foundation models, multimodal AI, embodied AI, AI for science, and AI safety.

Distribution channels include open-source distribution through GitHub and Hugging Face, published research output through major academic venues, the annual BAAI Conference, and continued cross-institution research-coordination relationships.

Benchmarks and standing

BAAI's evaluation framework focuses on foundation-model benchmarks (with the FlagEval benchmarking framework as the principal evaluation output), AI hardware-benchmarking metrics through FlagPerf, and published research output rather than direct commercial-product benchmarks. The Aquila series has been characterized in Chinese-language NLP industry coverage as a open-weights bilingual foundation model alternative.

The Wu Dao 2.0 1.75-trillion-parameter mixture-of-experts model briefly held the title of the largest published foundation model globally at the time of the June 2021 release. Subsequent releases of frontier-scale models from OpenAI, Google DeepMind, Anthropic, and other frontier AI labs have shifted the foundation-model scale landscape; the Wu Dao 2.0 weights have not been publicly released for direct benchmark comparison.

BAAI's open-source contributions through the FlagOpen ecosystem have continued to anchor the institute's open-research credibility. The FlagAI training framework, FlagPerf hardware benchmarks, FlagEval foundation-model evaluation, and other open-source projects have continued to receive contributions from the broader Chinese AI research community through 2024 to 2026.

Leadership

As of April 2026, BAAI's senior leadership includes:

• Zhongyuan Wang, Director (Chairman) of BAAI. Chinese AI research leader.
• Jie Tang, Chief Scientist of Wu Dao 2.0 and Tsinghua University professor. Chinese NLP and knowledge-graph researcher.
• Senior research leadership across the principal research programs (foundation models, multimodal AI, embodied AI, AI for science, AI safety).

Departures and arrivals are continuous. Bowen Zhou, the former director of BAAI's Foundation Model Research Center, departed to focus on Z.AI (Zhipu AI). Hang Li, former Bytedance chief scientist and a prominent NLP researcher, was associated with BAAI through 2023. The 2024 to 2026 period has continued senior research-talent recruitment alongside the Chinese AI commercial-startup spinout pattern.

Funding and backers

BAAI operates under Chinese government funding through the Beijing Municipal Government and the Ministry of Science and Technology, plus contributions from the founding member companies (Baidu, Alibaba, Tencent, ByteDance, Xiaomi, Megvii) and the founding academic institutions. Specific budget allocations have not been publicly disclosed, although industry coverage has placed BAAI's annual operating budget in the multi-hundred-million-yuan range.

The Chinese government commitment through the Beijing Municipal Government provides BAAI with financial-runway certainty. Open questions on near-term resourcing are limited, given the multi-year research-mandate commitments.

Industry position

BAAI occupies a structurally distinctive position as one of the principal Chinese AI research bodies, with Chinese government backing, the high-profile Wu Dao foundation model series, the FlagOpen open-source ecosystem, and the convening role through the annual BAAI Conference. The 2018 to 2020 founding mandate and the subsequent research output have positioned BAAI as one of the principal Chinese AI research bodies internationally.

Industry coverage has consistently characterized BAAI as one of the principal coordinating institutions for the Chinese AI research community, alongside the Shanghai AI Laboratory and Huawei Noah's Ark Lab on the industrial-research side. The 2024 to 2026 period has continued the institute's research output and the cross-institution research-coordination role.

Competitive landscape

• Shanghai AI Laboratory. Chinese government-backed AI research peer with a similar coordinating role for the Shanghai AI research ecosystem.
• Huawei Noah's Ark Lab. Chinese industrial AI research peer with overlap on foundation-model research.
• Z.AI (Zhipu AI), DeepSeek, Moonshot AI, Stepfun, MiniMax. Commercial Chinese AI startups with cross-institution research-cooperation relationships and talent flow from BAAI.
• Tencent Hunyuan, Alibaba Qwen / DAMO, Baidu, ByteDance Seed. Chinese commercial AI peers with member-institution backing of BAAI.
• Tsinghua KEG, Tsinghua IIIS. Chinese academic AI research peers with cross-institution research-cooperation.
• Allen Institute for AI (Ai2), EleutherAI, LAION. Non-Chinese open-research AI peers with overlap on the open-source AI infrastructure positioning.
• Hugging Face. Open-research distribution partner; BAAI distributes selected Aquila and FlagOpen variants through Hugging Face.

Outlook

• The continued cadence of FlagOpen ecosystem releases through 2026 and 2027.
• The next-generation Wu Dao or successor foundation model release timeline.
• The continued cross-institution research-coordination role across Chinese academic and industrial AI organizations.
• The continued talent flow from BAAI to commercial Chinese AI startups.
• The competitive dynamic with Shanghai AI Laboratory and Huawei Noah's Ark Lab on Chinese government-backed AI research positioning.
• The continued senior research-talent recruitment and senior leadership stability through the 2026 to 2027 Chinese AI commercial expansion.

Sources

• BAAI official site (English). Institute reference.
• FlagOpen ecosystem. Open-source AI infrastructure.
• BAAI on GitHub (FlagAI). Foundation-model training framework.
• Wu Dao 2.0 announcement coverage. Wu Dao foundation model series reference.
• BAAI Conference. Annual academic AI conference.