Intern S2 Preview 397B — Hardware Requirements & GPU Compatibility

Intern-S2-Preview-397B is InternLM's largest scientific multimodal foundation model, a 404.1-billion-parameter mixture-of-experts model (512 experts, 10 active per token, about 25.1 billion active parameters) positioned as the flagship above the smaller 36.1-billion-parameter Intern-S2-Preview. It introduces a vision-language pretraining paradigm that learns directly from raw pages of scientific literature rather than pre-parsed text, jointly modeling symbolic and visual relationships, and scales scientific reinforcement learning across more than 20 domains alongside long-horizon agent reinforcement learning in large sandboxed environments. InternLM reports leading general-reasoning results among open models plus strong biomolecular-interaction and material-structure generation performance. At its size it requires a multi-GPU server even quantized. Context length is 262,144 tokens, with text-reasoning benchmarks in the card evaluated at up to 256,000 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in July 2026.