# Shanghai AI Laboratory

> Source: https://aiwiki.ai/wiki/shanghai_ai_lab
> Updated: 2026-06-24
> Categories: Chinese AI, Research Organizations
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

Shanghai Artificial Intelligence Laboratory (Chinese: 上海人工智能实验室, commonly abbreviated Shanghai AI Lab or SHLAB) is a state-backed, non-profit AI research institution in Shanghai, China, that develops and openly releases frontier models under the "Scholar" or Shusheng (书生) brand, including the [InternLM](/wiki/internlm) large language model series and the [InternVL](/wiki/internvl) multimodal models. Unveiled in July 2020 at the World Artificial Intelligence Conference (WAIC), it operates as a "new-type research institution," a category of publicly supported, non-commercial research body in China, and is widely counted among the country's leading AI labs alongside its work on AI safety and large-model evaluation. [1][2][4] In addition to InternLM and InternVL, the lab is known for the InternImage vision backbone and the OpenMMLab and OpenCompass open ecosystems. [3][4]

## When was the Shanghai AI Laboratory founded?

The lab was inaugurated at WAIC in July 2020 and is based in the Xuhui District of Shanghai, at the West Bund media cluster on Longjing Road. [1][2] Its founding director was the computer scientist Tang Xiao'ou (汤晓鸥), who was also a professor at the Chinese University of Hong Kong (CUHK) and a co-founder of [SenseTime](/wiki/sensetime). [5][6] The lab built strategic ties with universities including Shanghai Jiao Tong University, Fudan University, Zhejiang University, and CUHK. [2]

The lab describes its mission as conducting strategic, original, and forward-looking research aimed at breakthroughs in fundamental AI theory and core technologies, and at building a large-scale, integrated research base in support of China's AI sector. [2] Stated focus areas include scientific multimodal foundation models, embodied AI, AI safety, large-model evaluation, and open data platforms. [2] In external profiles the lab is characterized as one of China's leading AI research institutions and as an organization that pairs frontier model development with technical work on AI safety. [4] At WAIC 2024 its director argued for an "urgent rebalancing of AI investment toward safety research," comparing AGI to nuclear fusion and describing safety as "a global public good." [4][7]

## What is the Scholar (Shusheng) model family?

The Shusheng (书生) program is the umbrella brand for the lab's foundation models. Its name is usually rendered "Scholar" in English; the language models specifically carry the sub-name Puyu (浦语), giving the full Chinese title 书生·浦语 for InternLM. [3][8]

| Project | First release | Notes |
| --- | --- | --- |
| InternImage | 2022 (CVPR 2023 Highlight) | Vision backbone using deformable convolutions; released under OpenGVLab. [9] |
| InternVL | December 2023 (CVPR 2024 Oral) | Vision-language model scaling the vision encoder to about 6B parameters. [10][11] |
| InternLM-7B | July 2023 | First open InternLM language model; weights open for research and free commercial use. [8] |
| InternLM-20B | September 2023 | 20B model, 60 layers, trained on over 2.3T tokens. [12] |
| InternLM2 | March 2024 | 6-dimension, 30-benchmark evaluation; long-context to 200K tokens. [13] |
| InternLM2.5 | July 2024 | 1.8B / 7B / 20B variants; 7B-Chat-1M supports up to 1M-token context. [21] |
| InternLM3-8B | January 2025 | Trained on about 4T tokens; MATH-500 83.0, AIME2024 20.0 in thinking mode. [22] |
| Intern-S1 | July 2025 | Scientific multimodal MoE model, 241B total / 28B active parameters. [14][15] |

### How do InternLM and InternVL work?

[InternLM](/wiki/internlm) is the lab's flagship large language model line. An initial 104B-parameter research model trained on roughly 1.6T tokens was described in a 2023 technical report. [16] The first openly released checkpoint, InternLM-7B, appeared in July 2023, with its weights made available for academic research and free commercial use. [8] InternLM-20B followed in September 2023, developed jointly with SenseTime, CUHK, and Fudan University; it used a deeper 60-layer architecture and over 2.3T training tokens. [12] The InternLM2 technical report was published in March 2024, reporting gains across six capability dimensions and thirty benchmarks and strong long-context behavior on a 200K-token "needle-in-a-haystack" test. [13] InternLM2.5 arrived in July 2024 with native tool-use support, and its 7B-Chat-1M variant extended context to about 1 million tokens. [21]

InternLM3-8B-Instruct, released on 15 January 2025, was trained on roughly 4 trillion high-quality tokens. The project reports that this data-centric approach achieves "saving more than 75% of the training cost compared to other LLMs of similar scale," while the model scores 83.0 on MATH-500 and 20.0 on AIME2024 in its deep-thinking mode. [22]

[InternVL](/wiki/internvl) is the multimodal counterpart, developed primarily through the lab's [OpenGVLab](/wiki/opengvlab) group. The original InternVL paper, "InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks," scaled the vision foundation model to roughly 6 billion parameters and aligned it with a language model using web-scale image-text data; it was presented as an oral at CVPR 2024. [10][11] Follow-up releases such as InternVL-Chat-V1.5 (April 2024) were positioned as open alternatives approaching the performance of contemporary proprietary multimodal systems on benchmarks like MMMU, DocVQA, and MathVista. [11] InternImage, an earlier [computer vision](/wiki/computer_vision) backbone built on deformable convolutions, was a CVPR 2023 Highlight and integrates with the OpenMMLab detection and segmentation toolkits. [9]

### What is Intern-S1?

In July 2025 the lab released and open-sourced Intern-S1, described as "the first open-source general-purpose model to integrate advanced scientific capabilities," at the 2025 World Artificial Intelligence Conference. [14][23] It uses a Mixture-of-Experts architecture with 241 billion total parameters and about 28 billion activated parameters, built on a Qwen3-235B language core, and was trained on roughly 5T tokens with more than 2.5T drawn from scientific domains. [14][15] The model pioneers a "cross-modal scientific reasoning engine" for interpreting data such as chemical formulas, protein structures, and seismic-wave signals, and the lab reports that on specialized scientific tasks it "outperformed leading closed-source models like Grok-4" while leading open multimodal models such as InternVL3 and Qwen2.5-VL on overall cross-modal capability. [14][23] A lightweight 8B variant, Intern-S1-mini, built on Qwen3-8B, followed in September 2025. [15]

## Is the Shanghai AI Laboratory open source?

A defining feature of the lab is the breadth of its [open source](/wiki/open_source) releases, distributed across GitHub organizations including InternLM, OpenGVLab, OpenMMLab, and OpenCompass, and on the Hugging Face hub. [3][4]

OpenMMLab is a family of PyTorch computer vision toolkits whose first components were released in October 2018 by researchers from CUHK and SenseTime, and which is now primarily maintained in association with the lab; Lin Dahua is the project's lead. [17] Its best-known toolboxes are MMDetection, an object detection benchmark and toolbox, and MMSegmentation for semantic segmentation, alongside foundational libraries such as MMCV and MMEngine. [17][18] OpenGVLab is the group behind the Intern vision and multimodal models, including InternImage and InternVL. [9][11]

The lab also maintains a wider toolchain around its models, including the XTuner training and fine-tuning framework, the LMDeploy inference and deployment framework, the MinerU document-parsing tool, and the MindSearch search application. [3]

## How does OpenCompass evaluate large models?

OpenCompass is an open-source [large language model](/wiki/large_language_model) evaluation platform established by the lab. [19] It supports a wide range of open-weight and API models, including Llama 3, Mistral, Qwen, GLM, GPT-4, Claude, and the lab's own InternLM, across more than 100 datasets, and provides leaderboards such as CompassRank. [19] OpenCompass 2.0 is organized around three components, CompassKit, CompassHub, and CompassRank, the last of which combines open and proprietary benchmarks; the project publishes a periodic leaderboard covering language, knowledge, reasoning, mathematics, coding, instruction following, and agent capabilities. [19] A companion toolkit, VLMEvalKit, extends similar evaluation to large multimodal models. [19]

## Who leads the Shanghai AI Laboratory?

Tang Xiao'ou (Sean Tang, 1968 to 2023) was the founding director of the lab. He was a professor in the Department of Information Engineering at CUHK, a co-founder of [SenseTime](/wiki/sensetime), and is widely associated with the development of deep-learning-based face recognition. [5][6] He died on the night of 15 December 2023 at the age of 55, after an illness that SenseTime did not disclose, and a memorial service was held in Shanghai on 19 December 2023. [5][20]

Following Tang's death, Zhou Bowen (周伯文) became director of the lab, making his public debut in the role at WAIC on 5 July 2024. [7] Zhou previously spent more than a decade at IBM, where he served as chief scientist of the Watson Group and director of the AI Foundations Labs, and later led AI research at the e-commerce company JD.com. [7] At WAIC 2024 he set out an "AI-45 degree Law," proposing that AI safety should advance in step with AI capability rather than lag behind it. [4] Yu Qiao (Qiao Yu) is identified as a leading scientist at the lab, associated with its open-source image and video generation work, and Lin Dahua is closely connected to the lab and leads the OpenMMLab project. [4][17]

## How is the Shanghai AI Laboratory funded?

The lab is described as a state-backed, new-type research institution rather than a commercial company, and is counted among China's "new research and development institutes." [1][7] External analysts characterize state-backed labs like the Shanghai AI Lab as "frontrunners in advanced AI development, and increasingly in governance," and note its role in technical AI safety work and in Chinese standards efforts for large models. [4] Its founding, leadership, and major projects are documented in the lab's own materials, the InternLM and InternVL technical reports, and contemporary technology press; precise figures for its budget and headcount are not publicly itemized in the sources cited here, and are therefore not stated. [2][4]

The Shanghai AI Laboratory should not be confused with the Beijing Academy of Artificial Intelligence ([BAAI](/wiki/baai)), a separate Chinese research institute that produces the BGE text-embedding models; for broader context on China's research landscape see [China AI](/wiki/china_ai). [4]

## References

1. "Former IBM scientist helms Shanghai AI Lab after death of SenseTime's Tang Xiao'ou," South China Morning Post. https://www.scmp.com/tech/tech-trends/article/3269319/former-ibm-scientist-helms-shanghai-ai-lab-after-death-sensetimes-tang-xiaoou
2. "About Us," Shanghai Artificial Intelligence Laboratory. https://www.shlab.org.cn/aboutus
3. "InternLM (Official release of InternLM series)," GitHub. https://github.com/InternLM/InternLM
4. "Profile: Shanghai AI Lab: Driving both AI safety and development," MERICS. https://merics.org/en/comment/profile-shanghai-ai-lab-driving-both-ai-safety-and-development
5. "SenseTime's founder Tang Xiao'ou dies at age 55, AI company says," South China Morning Post. https://www.scmp.com/tech/big-tech/article/3245325/tang-xiaoou-cuhk-professor-and-founder-ai-giant-sensetime-dies-age-55
6. "In Memory of Prof. Xiaoou Tang (1968-2023)," CUHK Information Engineering. https://www.ie.cuhk.edu.hk/in-memory-of-prof-xiaoou-tang-1968-2023/
7. "Former IBM, JD.com AI Expert Now Leads China's Shanghai AI Lab," Tech Times. https://www.techtimes.com/articles/306406/20240705/former-ibm-jd-com-ai-expert-now-leads-china-shanghai.htm
8. "internlm/internlm-7b," Hugging Face. https://huggingface.co/internlm/internlm-7b
9. "OpenGVLab/InternImage," GitHub. https://github.com/OpenGVLab/InternImage
10. "InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks," arXiv. https://arxiv.org/abs/2312.14238
11. "OpenGVLab/InternVL," GitHub. https://github.com/OpenGVLab/InternVL
12. "internlm/internlm-20b," Hugging Face. https://huggingface.co/internlm/internlm-20b
13. "InternLM2 Technical Report," arXiv. https://arxiv.org/abs/2403.17297
14. "Intern-S1: A Scientific Multimodal Foundation Model," arXiv. https://arxiv.org/pdf/2508.15763
15. "Release of Lightweight Scientific Multimodal Model Intern-S1-Mini," Oreate AI Blog. https://www.oreateai.com/blog/release-of-lightweight-scientific-multimodal-model-interns1mini-achieving-a-balance-between-general-and-specialized-capabilities-with-8b-parameters/cd2b089e195fb316705a8b8a81fb7e4c
16. "InternLM-techreport," GitHub. https://github.com/InternLM/InternLM-techreport
17. "OpenMMLab," GitHub. https://github.com/open-mmlab
18. "open-mmlab/mmdetection: OpenMMLab Detection Toolbox and Benchmark," GitHub. https://github.com/open-mmlab/mmdetection
19. "open-compass/opencompass," GitHub. https://github.com/open-compass/opencompass
20. "Funeral and Memorial Service for Professor Tang Xiao'ou," SenseTime. https://www.sensetime.com/en/news-detail/51167405?categoryId=1072
21. "InternLM2.5 release notes," GitHub. https://github.com/InternLM/InternLM/blob/main/README.md
22. "InternLM3-8B-Instruct," GitHub. https://github.com/InternLM/InternLM
23. "Shanghai AI Lab Releases and Open-Sources the First Multimodal Foundation Model with Scientific Reasoning," TMTPost. https://en.tmtpost.com/news/7641214