InternLM has introduced Intern-S2-397B, a 397-billion parameter multimodal foundation model designed for scientific intelligence and long-horizon agents. The model scales across pre-training, reinforcement-learning task coverage, and interactive agent environments to enhance general reasoning and agentic capabilities.
- New Pre-training Paradigm: Intern-S2-397B learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships without intermediate parsing to preserve text-visual correspondence.
- Scientific Modality Reasoning: By scaling diverse scientific reinforcement-learning tasks across more than 20 domains, the model achieves leading general-reasoning performance among open-source models.
- Long-Horizon Agents: The model connects multiple agent frameworks to large-scale sandboxed environments for black-box agentic reinforcement learning, improving generalization in long-horizon tasks.
This release aims to deliver a step change in scientific problem solving and the capability ceiling for long-horizon tasks in both general and scientific domains.