Intern-S2-397B

The full-scale sibling to the efficient 36B Intern-S2-Preview: a 397B-A17B MoE multimodal scientific foundation model (image + text + time-series), RL-trained across 20+ scientific domains with capabilities including biomolecular design and materials generation. 256K text context / 64K multimodal; BF16 with an FP8 sibling. Shanghai AI Lab's heavyweight AI-for-science model, staged for WAIC.

The non-preview Intern-S2-397B weights shipped on 13 September 2026 under Apache 2.0, described by the lab as its most capable multimodal foundation model for scientific intelligence and long-horizon agents: visual pre-training directly on raw pages of scientific literature, jointly scaled reinforcement-learning tasks across more than 20 scientific domains, and black-box agentic RL in large sandboxed environments, evaluated with OpenCompass, VLMEvalKit, and AgentCompass. Provenance: the config is Qwen3.5-397B-A17B's (60 layers, 512 experts, 10 active) with an extended vocabulary, so the entry is a continued-pretraining derivative rather than a from-scratch model.