397B-to-403B class multimodal model (Hugging Face lists approximately 403B parameters)
256K tokens text / 64K tokens multimodal evaluation
Apache-2.0
Official large-scale successor in the Intern-S2 scientific multimodal family from Shanghai AI Laboratory and the InternLM team, positioned as the project’s most capable open-weight model for scientific intelligence and long-horizon agents. Intern-S2-Preview-397B scales the earlier recipe through a vision-language pre-training paradigm based on raw rendered scientific pages, large-scale multi-task reinforcement learning across more than 20 scientific domains, and long-horizon agent reinforcement learning in sandboxed interactive environments. The full technical report documents scientific multimodal pre-training followed by supervised fine-tuning, scalable multi-task RL, black- and white-box agentic RL, and on-policy distillation, with training-system techniques including partial rollouts with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly. The model learns directly from rendered scientific literature while also using layout-aware interleaved PDF data and large-scale scientific-image retrieval, jointly modeling text, equations, tables, figures, page structure, and other visual evidence. Its upgraded time-series encoder supports approximately 300,000 time steps and is reported to run five to six times faster at maximum length while using around 20% of the previous implementation’s GPU memory; a separate forecasting branch adds direct numerical time-series generation. The separately released Intern-MemDec-4B provides a modular biological-memory extension without modifying the frozen 397B backbone: both branches process the same context, and a lightweight token-level router dynamically fuses their next-token distributions. InternLM reports that this attachment raises the average Biology-Instructions score from 56.92 to 60.32 while broadly preserving general and multimodal capability. Release materials also emphasize native tool calling, thinking and non-thinking modes, scientific generation, and integration with OpenClaw, Hermes, Claude Code, Transformers, vLLM, SGLang, and LMDeploy.