Shanghai AI Lab profile, products and governance

Anyone evaluating open-source Chinese AI research should know Shanghai AI Lab, not a startup chasing an IPO but a state institute whose director, Bowen Zhou, also chairs China's national AI safety standards group. Its models ship free under the MIT license with no hosted API tier, so using them means self-hosting or a third-party inference provider.

Shanghai AI Lab is a government-backed Chinese research institute founded in 2020 that builds open-weight foundation models, most notably the InternLM and Intern-S1 families and, since September 2026, the agentic Atria Dawn Preview model. It differs from private labs by publishing safety benchmarks Chinese regulators use to set AI standards.

Founded: 2020 · HQ: Shanghai, China · Team: 201-500 · CEO: Bowen Zhou · Funding: Government-funded via Shanghai municipal and central government research grant cycles; not venture-backed, no disclosed budget or funding round.

About Shanghai AI Lab

Shanghai Artificial Intelligence Laboratory, known publicly as Shanghai AI Lab, was unveiled at the World Artificial Intelligence Conference in Shanghai in July 2020 as a national level new type research institute backed by the Shanghai municipal government. It was established to give China a state funded counterpart to venture backed labs such as OpenAI and Google DeepMind, with a mandate to run foundational AI research spanning machine learning, computer vision and cross disciplinary scientific applications rather than build a single consumer product. The institute is headquartered in the Xuhui District of Shanghai. Its founding leadership included Xiaoou Tang, Andrew Yao and Jie Chen; after Tang's death, Bowen Zhou, a former IBM Watson Group chief scientist and JD.com cloud and AI president, took over as Director and Chief Scientist in July 2024.

The lab's flagship output is the InternLM family of large language models, now in its third generation, released as fully open weight checkpoints under its InternLM organization on GitHub and Hugging Face. Alongside InternLM it publishes InternVL, a multimodal vision language series whose 3.5 release in 2025 the lab says matched or beat GPT-5 on open source multimodal perception evaluations, and Intern-S1, a multimodal scientific reasoning model that scaled to a 1 trillion parameter mixture of experts version, Intern-S1-Pro, open sourced in early 2026. It also maintains the open source tool chain researchers use to build and evaluate these models: the OpenCompass evaluation framework, the LMDeploy and XTuner training and inference tools, and MinerU, a document parsing library that has drawn more than 48,000 GitHub stars.

In September 2026 the lab shipped Atria Dawn Preview, a 744 billion parameter mixture of experts agentic model built through post training on top of Zhipu AI's GLM-5.2 base, aimed at long horizon research and engineering tasks such as literature driven experiment design, code implementation and vulnerability analysis. The release came from ATRIA, a cross institutional research initiative the lab runs with Fudan University's Natural Language Processing Lab, led by Fudan associate professor Tao Gui, with Zhou himself among the roughly 140 listed authors on the accompanying arXiv paper. Earlier in 2026 the lab also opened China's first fully unmanned AI driven laboratory for materials science at Shanghai's Caohejing Hi-Tech Park, and at the 2026 World AI Conference it presented ten joint science results under its Shusheng Duan Yan research platform, including a protein design system, Shusheng AMix, and ChipDesign Agent, an autonomous electronic design automation tool the lab says cuts chip layout planning time by more than half.

Shanghai AI Lab is funded through Shanghai municipal and central government grant cycles rather than venture capital, putting it in the same state research institute category as Beijing's BAAI and Shenzhen's Pengcheng Laboratory. It has not disclosed an annual budget, revenue figure, or funding round, and unlike its venture backed peers it has no equity investors, valuation, or path to an IPO to report.

LinkedIn lists the institute at 201 to 500 employees, though that figure almost certainly understates its actual research headcount: alphaXiv alone tracks 184 individually credited researchers publishing under the lab's name, and the Atria paper alone lists roughly 140 co-authors drawn from partner universities. Zhou Bowen leads the lab as Director and Chief Scientist while holding a concurrent chair professorship in Tsinghua University's Department of Electronic Engineering, and he chairs the AI Safety Working Group, WG9, of China's national cybersecurity standards body TC260, giving the lab a direct hand in the country's official AI safety standards.

The lab states its mission as building open, controllable and trustworthy AGI by combining general purpose and domain specialized capabilities with open sharing and safety assurance. It backs that with a dedicated safety stack, Intern-SafeWork, and with published benchmarks that grade large language models on human value alignment, robustness to adversarial prompts and misuse risks such as toxic content generation. It has also taken part in the Global Workshop on the Safety of Open Source AI and co-signed a statement on international collaboration around open source AI safety, positioning itself as China's most visible institutional voice on AI safety standards rather than purely a model shop.

Within China's open weight model space, Shanghai AI Lab competes most directly with DeepSeek on frontier scale open source releases, though DeepSeek is privately backed by the hedge fund High-Flyer while Shanghai AI Lab is state funded and publishes far more of its research as academic papers and safety benchmarks. On the foundation model layer beneath Atria, it depends on Zhipu AI's GLM-5.2 base rather than training every flagship model from scratch, a dependency few Western labs of comparable output carry. Its core advantage over both is open, low friction distribution: every major release ships as downloadable weights under the MIT license with no waitlist or paid API tier, though that same openness means it captures no direct revenue from usage the way a hosted API business does.

As a state funded national institute operating inside China's regulatory system, Shanghai AI Lab sits on the compliance and standards setting side of Chinese AI policy rather than merely a company answering to it: Director Zhou Bowen chairs TC260's AI safety standards working group, and the lab is a named participant in China's state AI safety institute discussions tracked by Western policy researchers such as Carnegie and Stanford's DigiChina. It has not published an EU AI Act classification, GDPR compliance statement, or Western style trust center, reflecting that its primary regulatory relationship is with Chinese rather than international authorities.

Mission

Open, controllable and trustworthy AGI, built through open research, safety assurance and integration of general and domain-specialized AI capabilities.

Products

• InternLM (Open-weight LLM series): https://github.com/InternLM/InternLM
• Intern-S1 / Intern-S1-Pro (Multimodal scientific reasoning model): https://huggingface.co/internlm
• InternVL (Multimodal vision-language model): https://github.com/OpenGVLab/InternVL
• OpenCompass (Open-source LLM evaluation framework): https://github.com/open-compass/opencompass
• Atria Dawn Preview (Agentic foundation model): https://github.com/atria-asi/Atria-Dawn-Preview

Links

• Website
• GitHub
• Twitter
• LinkedIn
• Docs