A researcher wants a model to read a paper, check a crystal structure, work through a proof, and remember what it concluded three steps ago. A chatbot that is excellent at conversation can still fail at that, because the hard part is not fluency. It is holding a long chain of specialised reasoning together without drifting.
That is the problem Shanghai AI Laboratory (上海人工智能实验室) set itself with Intern-S2 (书生-S2), a science-oriented open model released at the 2026 Pujiang Innovation Forum in September. The lab calls its previous generation Intern-S1. The new model, published under the name Intern-S2-397B, is built around a simple structural idea: separate general reasoning from specialist knowledge, so that one can be updated without disturbing the other.
Key takeaways
- Intern-S2 has 397 billion total parameters and was released as an open model in September 2026.
- Its architecture splits a general reasoning base from a plug-in memory module that stores domain knowledge.
- The model is trained and optimised for Ascend (昇腾) domestic accelerators as well as open inference stacks.
- It is designed for long-horizon scientific work: reading papers, interpreting molecular and crystal structures, and multi-step mathematical reasoning.
- The lab reports strong results on science benchmarks; those figures come from the developer and need independent replication.
Why a separate memory module matters
Most large models (大模型) store what they know in the same weights that do the reasoning. Teaching them a new specialty means fine-tuning, which risks degrading capabilities they already had. For a research institution serving many disciplines, that is an awkward trade-off.
Intern-S2 takes a different route. According to the lab's description, a reasoning base handles long context, code, and mathematical proof, while a separate module holds subject-specific knowledge. Adding a new field means extending the memory module rather than retraining the base. The lab describes this as a knowledge-and-reasoning separation design; an outside analogy would be a librarian who stays the same while the shelves are restocked.
The practical claim is that the model can keep up with new scientific literature without forgetting how to think. Whether that holds outside the lab's own evaluation is a separate question, and one that open weights make possible to test.
What it is being asked to do
The release targets tasks that look unlike ordinary chat. The lab points to Biology-Instructions and Mol-Instructions, which test understanding and generation of biological and molecular content, including material structure. It also cites IMO-Proof and AdvancedMathBench, benchmarks built around competition-level mathematics and long proofs.
Long-horizon work is the common thread. A proof that runs for several pages, or a literature review that must stay consistent across dozens of sources, punishes models that lose track of earlier steps. Intern-S2 is also being wired into the lab's scientific discovery platform, Intern-Duanyan (书生·端砚), which was upgraded alongside the model release. The platform is intended to connect model output to large scientific facilities, including the Shanghai Synchrotron Radiation Facility and the China Spallation Neutron Source.
That link is the more interesting part of the announcement. A model that proposes an experiment is useful; a platform that can send the proposal to a beamline, collect data, and feed the result back into the next iteration is a different category of tool. The lab says the platform now includes an evidence chain meant to make each step traceable, and connections to pilot-scale testing facilities.
The domestic compute angle
Intern-S2 was optimised together with the Ascend ecosystem, covering computation, communication, and memory use. For Chinese research institutions that cannot assume access to the latest Nvidia hardware, a model whose weights and training recipe work on domestic accelerators has value beyond benchmark scores.
The same model is available through several open repositories, and the lab offers a free API quota with an application route for higher limits. That distribution choice is consistent with how Chinese public research institutions have built ecosystems around open models: publish the weights, document the training, and let universities adapt them.
Honest limitations
- The benchmark claims come from Shanghai AI Laboratory's own release. I found no independent reproduction of the reported science results.
- Benchmark performance does not establish usefulness in a real laboratory. Error rates on a live experiment are a different measurement.
- Details about the training corpus and the exact mechanism of the memory module are described at a high level in the announcement; readers should consult the model card and technical report before treating the architecture as settled.
- The connection to large scientific facilities is an integration roadmap described by the lab. The extent of routine operational use was not documented in the sources reviewed.
- This article relies on the lab's official announcement and People's Daily reporting, which largely restates the same release. English-language independent technical coverage was limited at the time of writing.
What readers can do now
- If you work in a scientific field, download the model card before the marketing summary. Check which benchmarks match your actual task.
- Try a small, reproducible test of your own: give the model a paper in your field and ask it to extract a specific method, then verify every claim against the source.
- For teams in China, test deployment on your own hardware stack. A model advertised as accelerator-friendly still needs a local benchmark run.
