35B-A3B-class MoE; 36B total parameters in Hugging Face metadata
256K tokens native; evaluations capped at 128K text / 64K multimodal
Apache-2.0
Official open-weight scientific multimodal foundation model from the InternLM team, positioned as an efficient successor for scientific reasoning, scientific generation, and workflow automation. Intern-S2-Preview-35B is a roughly 35B-total, 3B-active Mixture-of-Experts model continued-pretrained from Qwen3.5-35B. Its training emphasizes task scaling across hundreds of professional scientific tasks rather than parameter scaling alone, aiming to match the trillion-scale Intern-S1-Pro on multiple core scientific tasks with far fewer active parameters. The later Intern-S2 technical report documents a unified pipeline spanning scientific multimodal pre-training, supervised fine-tuning, scalable multi-task reinforcement learning, black- and white-box agentic RL, and on-policy distillation, supported by partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware agent-experience assembly. Release materials emphasize stronger multimodal scientific reasoning, improved scientific-agent capabilities, enhanced spatial modeling for small-molecule structures, and a real-valued prediction module; the team also describes it as the first open-source model combining material crystal-structure generation with strong general capabilities. Shared-weight Multi-Token Prediction with KL loss and chain-of-thought compression are used to improve generation speed and reasoning brevity while preserving capability. The model supports native tool calling and deployment through LMDeploy, vLLM, and SGLang. Official API documentation lists a 256K context window, while published evaluations cap text reasoning at 128K and multimodal evaluation at 64K.