SII

The Shanghai Innovation Institute (SII) (上海创智学院, est. September 2024) is a high-level innovation hub bridging academic research and industrial application, founded on the premise that "AI belongs to young people." Received ~500M yuan (~$70M) in donations with 50+ industry partners. Leadership: Executive VP Ding Xiaodong (丁晓东, formerly President of the University of Shanghai for Science and Technology), with VPs Wu Libo (Fudan), Qiao Yu, and Lu Cewu (SJTU). Its primary AI research arm is the GAIR Lab (Generative AI Research, GitHub, HuggingFace), directed by Pengfei Liu (Associate Professor, SJTU). Deeply integrated with Shanghai Jiao Tong University and the related SAIRI (Shanghai AI Research Institute, est. 2019, led by Song Haitao), with frequent collaborations with Ant Group/InclusionAI, Sand.ai, and Fudan NLP (OpenMOSS).

SII's research labs practice "radical openness" — releasing not just weights but full training trajectories (logs, intermediate checkpoints, data mixtures). The daVinci series spans pretraining science (daVinci-LLM-3B, 8T tokens, matches OLMo-3 7B at half the parameters), software engineering (daVinci-Dev 32B/72B, 58.5% SWE-Bench), agentic AI (daVinci-Agency, 353B, 47% gain from only 239 samples), and multimodal (daVinci-MagiHuman, 15B single-stream video+audio with Sand.ai).

The LIMO ("Less Is More") line demonstrated that 817 curated samples can match 100x more data (AIME24 63.3%, COLM 2025). Data Darwinism established a systematic L0–L9 taxonomy for data curation producing the Darwin-CC corpus (504B tokens, 1B+ HuggingFace downloads). ASI-Evolve explores whether AI can accelerate its own development, discovering 105 novel attention architectures surpassing recent benchmarks.