Intern-S2-Preview: Scientific Agentic Foundation Model
Intern-S2-Preview is pitched as a 397B scientific agent model for multimodal reasoning, tool use, and long-horizon tasks.
The paper describes a training stack built from scientific multimodal pre-training, supervised fine-tuning, multi-task RL, agentic RL, and on-policy distillation. It adds time-series modules for scientific signal understanding and forecasting, plus a separate Memory Decoder path for specialization without changing the frozen 397B backbone. The authors report competitive or leading benchmark results across scientific, multimodal, agentic, and general-purpose evaluations. ArXiv · AI/CL/LG's note
The paper describes a training stack built from scientific multimodal pre-training, supervised fine-tuning, multi-task RL, agentic RL, and on-policy distillation. It adds time-series modules for scientific signal understanding and forecasting, plus a separate Memory Decoder path for specialization without changing the frozen 397B backbone. The authors report competitive or leading benchmark results across scientific, multimodal, agentic, and general-purpose evaluations. ArXiv · AI/CL/LG's note
score 7