Intern-S2-Preview: Scientific Agentic Foundation Model
Intern-S2-Preview is a 397B scientific agent model with added time-series and memory-specialization paths.
The paper describes a multimodal training pipeline built on rendered scientific documents, image-text data, and scientific corpora. Post-training combines supervised tuning, multi-task RL, agentic RL, and on-policy distillation for longer scientific tasks. The authors say the 397B model is competitive or leading across scientific, multimodal, agentic, and general benchmarks. A separate 4B Memory Decoder raised the Biology-Instructions average score from 56.92 to 60.32 without changing the frozen 397B backbone. HF Daily Papers' note
The paper describes a multimodal training pipeline built on rendered scientific documents, image-text data, and scientific corpora. Post-training combines supervised tuning, multi-task RL, agentic RL, and on-policy distillation for longer scientific tasks. The authors say the 397B model is competitive or leading across scientific, multimodal, agentic, and general benchmarks. A separate 4B Memory Decoder raised the Biology-Instructions average score from 56.92 to 60.32 without changing the frozen 397B backbone. HF Daily Papers' note
score 6