Researchers behind Shanghai AI Laboratory's Intern model family have published Intern-S2-Preview, a series of foundation models built specifically for scientific work that requires reasoning over evidence in several modalities, operating scientific tools and environments, and sustaining progress across long task horizons.
The flagship, Intern-S2-Preview-397B, carries 397 billion parameters. The paper, posted to arXiv on August 13, lists more than 120 authors — an indication of the scale of the engineering effort behind it.
A pipeline built for long horizons
Training begins with scientific multimodal pretraining over rendered scientific documents, interleaved image-text data and broad scientific corpora. From that pretrained checkpoint the team applies a unified post-training pipeline: supervised fine-tuning, scalable multi-task reinforcement learning, both black-box and white-box agentic RL, and on-policy distillation.
Much of the paper's contribution lies in making that pipeline stable at scale. The authors describe partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization and trace-aware experience assembly for agentic tasks — a set of techniques aimed squarely at the rollout and training instabilities that make long-horizon reinforcement learning expensive.
Two architectural bets
Intern-S2-Preview-397B extends time series modelling beyond efficient long-sequence understanding into numerical forecasting, which matters for domains where the raw evidence is a signal rather than a document or an image. The team reports that these modules improve scientific signal understanding and forecasting on the SciTS benchmark.
The second bet is Memory Decoder, studied as a separate memory-augmented path rather than a change to the main model. The goal is rapid scientific specialization without touching the frozen 397B backbone. The team reports that a separate four-billion-parameter extension, Intern-MemDec-4B, lifted the average Biology-Instructions score from 56.92 to 60.32 — a gain achieved without modifying the underlying model at all.
That decoupling is arguably the most practically interesting result in the paper. Specializing a frontier-scale model for a narrow scientific domain normally means fine-tuning it, which is costly and produces a separate copy for every field. A small attachable module that delivers a comparable lift leaves the base model untouched and shareable across domains.
Where it lands
Across scientific, multimodal, agentic and general-purpose benchmarks, the authors report competitive or leading results in multiple settings — careful phrasing that stops short of claiming a clean sweep. The model is labelled a preview, and the paper reads as a progress report on an architecture still being explored rather than a finished product.
The framing is the notable part. Most frontier model releases are judged on coding, general reasoning and consumer-facing chat quality. Intern-S2-Preview aims at a narrower target: systems that read scientific evidence in several modalities at once, call scientific tools, and keep working through tasks that unfold over many steps rather than a single exchange.
That target has become a crowded ambition. Laboratories across the industry have been positioning agentic models as instruments for research itself rather than assistants to researchers. What separates the proposals is less the goal than the mechanics — how training stays stable across long agentic rollouts, and how a general model gets specialized without being rebuilt. On both counts, Intern-S2-Preview offers a concrete account.
The abstract does not specify release timing or licensing terms for the model weights, and the preview label suggests the series is still evolving. For researchers weighing whether to look closer, the numbers that carry the most information are the SciTS forecasting gains and the Biology-Instructions jump, because both test the architecture's specific claims rather than general capability.
