Intern-S2-Preview Released

InternLM has released the 35B scientific multimodal foundation model Intern-S2-Preview.

It was continually pre-trained based on Qwen3.5, applying full-chain training and task scaling that connects pre-training through to reinforcement learning (RL). The company explained that it scaled up hundreds of specialized scientific tasks to jointly improve general reasoning, multimodal understanding, and agent capabilities.

• It achieved performance on core scientific tasks comparable to the trillion-scale Intern-S1-Pro.
• It strengthened spatial modeling of small molecule structures and added a real-valued prediction module.
• It was presented as the first open-source model to combine material crystal structure generation with general capabilities.
• In the RL stage, shared-weight MTP + KL loss was used to reduce training-inference mismatch, and CoT compression shortened responses while improving both reasoning performance and speed.
• The company stated that it also showed stronger agentic ability than the previous generation on scientific agent benchmarks.