The Intern team described the scientific agentic model Intern-S2-Preview-397B for multimodal scientific reasoning

The Intern team published a description of Intern-S2-Preview on arXiv (2608.13505v1), a series of scientific agentic foundation models designed for multimodal scientific understanding, reasoning, generation, and solving long-horizon tasks. The training process begins with multimodal pre-training on rendered scientific documents, interleaved image-text data, and diverse scientific corpora. This is followed by unified post-training comprising supervised fine-tuning, scalable multi-task reinforcement learning, black-box and white-box agentic RL, and on-policy distillation. According to the authors, this process is complemented by techniques for stable and efficient training and rollouts — partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly for agentic tasks.

The Intern-S2-Preview-397B model has 397 billion parameters. According to the authors, it extends time-series modeling from effective understanding of long sequences to numerical prediction as well. The Memory Decoder module is studied separately as a memory-augmented route to rapid scientific specialization without modifying the frozen 397B base model — this variant is designated Intern-MemDec-4B.

According to the authors, Intern-S2-Preview-397B achieves competitive to leading results across scientific, multimodal, agentic, and general benchmarks. They report that the time-series modules improve scientific signal understanding and prediction on the SciTS benchmark. The Intern-MemDec-4B extension improves the average score on the Biology-Instructions benchmark from 56.92 to 60.32 without modifying the frozen 397B base model.

The source text is an excerpt from the abstract of a scientific publication and contains no information about weight availability, licensing, or pricing. Details can be found in the source article.