Shanghai AI Lab published an August 13 paper on its 397B scientific-agent model
Intern-S2-Preview combines visual pretraining, tool use and long-horizon reinforcement learning in an open 397B-parameter scientific model.
By RuntimeWire Staff · Published
Primary source: Aligned News - AI Intelligence
Why it matters
Intern-S2 packages page-level document understanding, scientific generation and agent workflows into open weights. Independent validation and reliable performance across long research runs remain the critical tests.
Kai Chen and Wenwei Zhang (@wenweiz97) are among more than 120 researchers who detailed Shanghai AI Lab's 397B-parameter bet on scientific agents in a technical paper published August 13. Intern-S2-Preview-397B is designed to work across scientific documents, images, time-series data and external tools while sustaining the longer chains of activity required for research workflows.
Shanghai AI Lab released the 397B model around July 16-18, 2026, according to the available research and the lab's API documentation. The technical paper arrived on August 13, providing a fuller account of a model that was already available rather than marking a same-day launch.
Chen heads Shanghai AI Lab's Large Model Center and previously led computer-vision research and development as a director at SenseTime. He earned an engineering degree from Tsinghua University and a Ph.D. from the Chinese University of Hong Kong. Zhang, a research scientist focused on post-training and AI agents, has led development of the MMDetection and MMDetection3D open-source projects. Their work fits within director and chief scientist Bowen Zhou's broader thesis that general models should retain broad capabilities while acquiring deeper expertise in individual fields.
A model that treats papers as visual objects
Intern-S2-Preview's most consequential design choice is its treatment of scientific literature. During visual pretraining, the model learns from rendered pages rather than relying entirely on text extracted from PDFs. The paper says this approach preserves figures, tables, equations and layout information that may be lost during text extraction, while linking visual units with their surrounding scientific context.
The model documentation says Intern-S2-Preview-397B was trained for scientific question answering, biomolecular interaction design, material-structure generation, coding and tool use. Shanghai AI Lab evaluated text reasoning with an inference limit of 256,000 tokens and multimodal tasks with a limit of 64,000 tokens. Those limits describe the evaluation setup, not proof that the model can complete an open-ended research program across either window without losing direction.
Shanghai AI Lab's training pipeline combines supervised fine-tuning, reinforcement learning across scientific domains, reinforcement learning inside interactive agent environments and on-policy distillation. A dedicated branch also handles numerical forecasting from time-series data, allowing the model to predict future values directly instead of expressing every number through text tokens.
That architecture reflects Shanghai AI Lab's view that scientific intelligence requires several kinds of work in one system: reading evidence, reasoning about it, generating scientific artifacts, calling tools and revising a plan after intermediate results. Static benchmark questions cover only the first part of that job.
Open weights at 397B scale
Shanghai AI Lab released the model under the Apache 2.0 license, with weights and deployment code available through Hugging Face and the InternLM GitHub repository. The naming is less tidy than the license. Shanghai AI Lab calls the model 397B, while Hugging Face lists approximately 404 billion parameters and about 810 GB of files.
Researchers can use Shanghai AI Lab's OpenAI-compatible API. The model's scale may narrow the group that can independently inspect and reproduce the full system. Universities and companies with suitable infrastructure can study the weights, serving code and model behavior. Smaller research groups will encounter a familiar frontier-model problem: the code is available, while meaningful experimentation still requires scarce compute.
The long-horizon claim
Shanghai AI Lab reports competitive or leading results across its collection of scientific, multimodal, agent and time-series benchmarks. Intern-S2-Preview-397B was compared with models including GPT-5-mini, Gemini 2.5 Flash, DeepSeek-V3 and Shanghai AI Lab's larger Intern-S1-Pro. Those results come from the Intern-S2 research team and have not been independently reproduced across the model's advertised range of tasks.
The paper is more careful than the model card's launch language. Its authors call Intern-S2-Preview a preview system and identify long-workflow reliability, stronger verifiers, deeper tool integration and additional domain specialization as remaining work.
Independent research offers a useful check. SciAgentArena, a 2026 benchmark covering about 200 real-world scientific tasks, found that current agents can contribute to well-specified data-analysis workflows. Performance remained uneven when tasks demanded novel insights, self-directed exploration or reliable answers to open-ended research questions. Intern-S2's benchmark scores therefore measure components of scientific work, not autonomous scientific discovery itself.
A different route to AI science
The market is splitting between model builders and companies assembling complete scientific systems. Edison Scientific's Kosmos combines literature review, data analysis and persistent research workflows. Lila Sciences is pairing AI reasoning with automated laboratories that can physically test generated hypotheses.
Shanghai AI Lab is taking a more horizontal route. The paper describes Intern-S2-Preview as a scientific agentic foundation model. A separate Memory Decoder can attach specialized knowledge to the frozen 397B backbone, an approach intended to let researchers add expertise without retraining the entire model.
Zhou framed scientific research as "the next Coding" in a Shanghai AI Lab presentation. The comparison explains the timing. Coding agents improved rapidly because programs provide fast, objective feedback. Scientific research brings slower experiments, ambiguous failures and questions whose answers may not exist yet.
Intern-S2 is the team's attempt to build for that messier loop. Its open release gives outside researchers a substantial system to examine, adapt and challenge. Whether it can keep making useful progress after the benchmark ends will matter more than another first-place table.