InternLM Releases Intern-S2-397B Multimodal Model for Scientific AI and Long-Horizon Agents

InternLM has launched Intern-S2-397B, a massive 397-billion parameter multimodal foundation model tailored for scientific intelligence and long-horizon interactive agents. The model introduces a novel pre-training paradigm that processes raw scientific paper pages directly to jointly model textual semantics and visual layout without intermediate text parsing. This release sets a new capability benchmark for open multimodal AI, enabling advanced scientific applications such as biomolecular interaction design and material structure generation. Furthermore, its integration with sandboxed environments for agentic reinforcement learning pushes forward the state of automated, multi-step scientific research workflows. Intern-S2-397B scales reinforcement learning across more than 20 specialized scientific domains to strengthen both general reasoning and specialized generation tasks. By learning directly in a shared representation space, it preserves text-visual correspondence, improves spatial reasoning, and significantly boosts data efficiency.