Shanghai AI Lab open-sources Intern-S1-Pro, a 1T-parameter science MoE model
Shanghai AI Lab just open-sourced Intern-S1-Pro, a massive 1-trillion-parameter multimodal model tailored for scientific research. Built on a 512-expert MoE architecture that only activates 22B parameters per token, it tackles complex problems across chemistry, physics, and biology at Olympiad-level precision. It is free for commercial use, offering researchers a powerful open base for AI-driven discovery.
Intern-S1-Pro is an open-source multimodal model developed by Shanghai AI Lab, designed specifically to assist researchers with chemical synthesis, materials science, genomics, and advanced scientific reasoning. Clocking in at 1 trillion parameters, it is currently the largest open-source AI model built for scientific discovery.
What it does
Intern-S1-Pro spans five core scientific disciplines: chemistry, materials science, biology, earth science, and physics, handling more than 100 specialized research workflows.
On the perception side, it parses complex molecular diagrams, scientific charts, and remote sensing imagery. For analytical tasks, it predicts physical and chemical properties, infers reaction conditions, and helps generate protein sequences.
In standard evaluations, the model demonstrated Olympiad-level precision on IMO-Answer-Bench for mathematics and IPhO2025 for physics. It also matched or outperformed top commercial models on multidomain benchmarks like SciReasoner, while showing strong interactive agent capabilities on the Tau-2 benchmark.
How it works and how to run it
Despite having 1T total parameters, Intern-S1-Pro uses a sparse Mixture-of-Experts (MoE) architecture with 512 experts. For any single token or query, it only activates 8 experts totaling 22B parameters, which keeps inference compute manageable.
To better understand physical signals, the researchers introduced Fourier Position Embedding (FoPE) and redesigned the sequence encoder. This allows the model to process both discrete text tokens and continuous wave data, spanning from microscopic biological telemetry to macroscopic astrophysical readings.
Training stability was tackled through dense routing estimation and grouped routing, ensuring balanced hardware loads across clusters. The lab co-developed the stack with Ascend and MetaX hardware ecosystems, utilizing the XTuner V1 framework for training and LMDeploy for inference.
Weights and code are available on GitHub, Hugging Face, and ModelScope under a license that permits free commercial use. An interactive web demo is accessible at chat.intern-ai.org.cn.
Why it matters
Scientific labs often struggle between paying for closed API models or settling for smaller, brittle fine-tuned checkpoints. By packing 1T parameters into an MoE that runs on 22B active compute, Intern-S1-Pro gives researchers serious Olympiad-tier reasoning on an open foundation.
While storing and serving 1T total weights still demands dedicated infrastructure, we think offering this scale under a free commercial license is a massive boost for open AI-driven research.
Sources (1)
- shanghaiopen.org.cn(opens in a new tab)Originalshanghaiopen.org.cn