The Shanghai AI Laboratory (Shanghai Artificial Intelligence Laboratory) is China's flagship state-backed AI research institute, founded in 2020 with SenseTime founder Tang Xiao'ou as founding dean. It functions as the academic anchor of the Chinese AI labs ecosystem: open models, shared evaluation infrastructure, and safety research, in the institutional role that AI2 and university labs play in the West.1
Model program
The InternLM line (from 2023) delivered fully documented open language models whose reports detail data pipelines and training dynamics; InternLM2's technical report became a standard reference for open pretraining practice.1 The InternVL series has been the consistently strongest open vision-language line from any institution, tracked head-to-head with commercial Gemini and GPT multimodal tiers on public suites, and the lab's media, scientific, and embodied-AI models extend the catalog. OpenCompass, its evaluation platform, is China's principal open benchmark harness, giving the lab agenda-setting influence over how domestic models are measured, MMLU-class suites included.2
Role
The laboratory convenes rather than competes: its researchers co-publish with Alibaba, universities, and startups; its safety and governance work feeds China's regulatory apparatus, including frontier-risk research aligned with the international dialogue documented on this wiki's governance pages; and its open releases seed the commercial ecosystem the way EleutherAI's and Hugging Face's artifacts do elsewhere. State funding insulates it from the consolidation pressures reshaping the startup tigers, and its Shusheng model brand, World AI Conference presence, and standards work on evaluation and frontier-safety practice keep it the sector's most visible non-commercial voice.23
Stated mission and record
The laboratory states its mission as conducting "strategic, original, and forward-looking scientific research and technological breakthroughs, breaking through important fundamental theories and key core technologies of artificial intelligence" (translated from Chinese).4 Its stated vision is to build "general artificial intelligence that fuses generality with specialization, is open and shared, and safe and trustworthy" (translated), organized around three pillars: generality-specialization fusion, open sharing, and safety and trustworthiness, the last framed through its "AI-45° Law" proposal for balancing capability and safety.5 Its Hugging Face organization adds, in English: "We keep open-sourcing high quality LLMs/MLLMs as well as a full-stack toolchain for development and application."6
As of July 2026, the open-sharing pillar is broadly consistent with practice: the InternLM and InternVL lines ship open weights with unusually detailed technical reports, and OpenCompass operates as a public evaluation harness. The safety pillar is reflected in frontier-risk research feeding China's governance apparatus. Two tensions sit alongside this record: the same mission statement casts the lab as "the source of China's original AI theory and key technologies" (translated), placing its openness inside a national-strategic mandate shared across Chinese AI labs; and its stewardship of OpenCompass means it both trains models and maintains a principal benchmark harness on which they are compared. Neither point is conceded or contested in the lab's own materials.42
See also
References
-
InternLM Team, "InternLM2 Technical Report," arXiv:2403.17297, March 2024. ↩↩
-
Chen, Z., et al., "InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks," arXiv:2312.14238, December 2023. ↩↩↩
-
Shanghai AI Laboratory institutional materials and coverage, 2020-2026. ↩
-
Shanghai AI Laboratory, "关于我们" (About Us), official website, as of July 2026; quotations translated from Chinese. ↩↩
-
Shanghai AI Laboratory, homepage, official website, as of July 2026; quotations translated from Chinese. ↩
-
InternLM (Shanghai AI Laboratory), organization self-description, official Hugging Face organization page, as of July 2026. ↩
