Shanghai AI Lab Releases Intern-S2-Preview-397B for Science
Shanghai AI Lab introduces Intern-S2-Preview-397B, a 397B multimodal model built for scientific reasoning, long-horizon agents, and research workflows.
What happened
Shanghai AI Lab introduced Intern-S2-Preview-397B, a 397-billion-parameter multimodal foundation model aimed at scientific intelligence and long-horizon agent work. The announcement appeared on the lab’s official account at 09:09 UTC on September 14, inside this edition’s window. The model is presented alongside a smaller 35B version and is available through the InternLM model ecosystem.
Intern-S2-Preview is designed to process scientific material in its original visual form rather than relying entirely on text extracted from papers. The lab says its training combines visual pre-training, reinforcement learning across more than 20 scientific domains, and agent environments that require extended interaction with tools and sandboxes. Claimed target capabilities include general reasoning, scientific problem solving, biomolecular interaction design, material-structure generation, and long-horizon task execution.
The release also matters because it is positioned as an open model rather than a closed research preview. The official repository lists Intern-S2-Preview-397B checkpoints for Hugging Face and ModelScope, Apache 2.0 licensing, and deployment paths through supported inference frameworks including LMDeploy. That gives research groups a route to inspect, adapt, and host the system, although the 397B scale makes practical deployment expensive and highly infrastructure-dependent.
Why it matters
The important shift is strategic: Chinese open-model work is moving from general chat benchmarks toward models trained around scientific workflows and tool use. The claims still require independent reproduction, especially for agent reliability and domain-specific results, but the combination of multimodal scientific data, large-scale reinforcement learning, and open checkpoints makes this more consequential than a routine model refresh.
Uncle Cat take
A 397B checkpoint is less interesting than the decision to train across 20-plus scientific domains; that breadth will determine whether Intern-S2-Preview becomes research infrastructure or just another leaderboard entry.