Intern-S2-Preview: 35B model rivals trillion-scale scientific AI

Shanghai AI Lab has released Intern-S2-Preview, an efficient 35B-parameter scientific multimodal foundation model that challenges conventional scaling wisdom. Instead of solely relying on more data or larger parameters, the team introduced task scaling—increasing the difficulty, diversity, and coverage of scientific tasks. The model is pretrained from Qwen3.5 and undergoes a full-chain training pipeline from pre-training to reinforcement learning. Despite its modest 35B parameters, it matches the performance of the trillion-scale Intern-S1-Pro on multiple core professional scientific tasks.

Intern-S2-Preview introduces several technical innovations. It strengthens spatial modeling for small-molecule structures and adds real-valued prediction modules, making it the first open-source model capable of material crystal structure generation while retaining strong general capabilities. For reinforcement learning, it uses shared-weight multi-token prediction (MTP) with KL loss to reduce training-inference mismatch, improving token generation speed and acceptance rate. It also applies chain-of-thought compression to shorten responses without sacrificing reasoning quality, achieving both performance and efficiency gains. The model also shows enhanced agent abilities on scientific benchmarks, making it a versatile tool for researchers.