# Intern-S1 and Intern-S2: a scientific model hub, not a code repo > The Intern-S1 repository holds announcements rather than code, with an English and a Simplified Chinese README, an assets directory and a docs directory, while the weights live on Hugging Face and ModelScope. What the READMEs describe is a family of multimodal scientific models, from the 1T-parameter MoE Intern-S1-Pro to a 35B preview built by task scaling rather than parameter scaling. **InternLM/Intern-S1** — A Scientific Multimodal Foundation Model - Repository: https://github.com/InternLM/Intern-S1 - Stars: 870 · Forks: 56 - Language: Unknown - License: Apache-2.0 - Published: 2026-10-20 · Updated: 2026-10-20 · Language: en - Canonical page: https://hysenlabs.com/projects/internlm-intern-s1 ## What is actually in the repository The top-level listing is short: a .gitignore, a .pre-commit-config.yaml, the Apache-2.0 LICENSE, a README.md, a README_zh-CN.md, an assets directory and a docs directory. That is the whole repository, and it is the first fact an engineer needs. There is no training code, no inference server, no configuration and no weights. The README itself points elsewhere for the things you would actually want: Hugging Face collections for Intern-S2 and Intern-S1, a ModelScope collection, technical reports on arXiv, and an online chat endpoint. The assets directory holds the banner and preview images the READMEs embed, and the docs directory is what the site is built from. So the repository is an announcement and documentation hub for a model family, maintained with a pre-commit hook, a bilingual README pair and a licence file, and the software itself lives in the weight repositories. The primary language field is empty in the repository metadata, which is consistent with a repository that contains no source code. The slug of the project is Intern-S1, and the technical report link for S1 is arXiv 2508.15763, but the current README leads with Intern-S2. The repository name and the current model family have drifted apart, which is worth knowing before you follow the links in either direction. ## S2-397B is described along three axes The introduction presents Intern-S2-397B as the most capable multimodal foundation model the project has for scientific intelligence and long-horizon agents, and it says the model scales along three critical dimensions: pre-training, reinforcement-learning task coverage, and interactive agent environments. The claim about the method is that combining a new vision-language pre-training paradigm with large-scale multi-task reinforcement learning and long-horizon agent reinforcement learning produces, in the README's words, a step change in general reasoning, scientific problem solving and agentic capabilities. The phrasing is the authors' own, and the README gives no benchmark table for it, so the evidence for the claim is the technical report and the model cards rather than this page. The three axes are a useful way to think about what the project is actually optimising, because they are not the same axes an ordinary language model scales. Pre-training is about how the model reads its data. Reinforcement-learning task coverage is about how many scientific task families the model is trained on. Interactive agent environments are about whether the model can act inside a loop with tools and a sandbox and learn from the outcome. A preview of the same model, Intern-S2-Preview-397B, is documented in a collapsed section of the README with the same three axes and the same feature list. Having the preview and the final model described identically in the same file is a small sign of how the announcement was assembled, and it is a reason to read the reports rather than the README. ## Reading raw pages of literature without parsing them first The first feature of the S2-397B is the most concrete idea in the README, and it is a data decision rather than an architecture one. Via visual pre-training, the model learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships in a shared representation space without intermediate parsing. The stated benefits follow from skipping the parsing step: it preserves text-visual correspondence, strengthens spatial and visual reasoning, and improves data efficiency. The reasoning is straightforward. A pipeline that parses a PDF page into text, extracts figures separately and stores them in another format throws away information at the moment of parsing, in particular the two-dimensional arrangement of a figure and the proximity of a caption to the thing it describes. Training on the rendered page keeps that arrangement inside the training signal, so a model can learn that a table's position and its header row are part of its meaning. For scientific documents this is not a marginal gain, because the conventions of a paper depend heavily on layout: a figure is not interchangeable with its caption, a table's axes matter, and a chemical structure drawn next to a paragraph is a different object from the same structure rendered as text. The cost of the approach is that the model needs a stronger visual backbone, and that the pretraining corpus has to be pages rather than extracted text, which is a rarer and more expensive kind of data. ## Scientific reinforcement learning across more than twenty domains The second feature concerns the range of tasks. By scaling diverse scientific reinforcement-learning tasks across more than 20 domains and training them jointly, the README says the model achieves leading general-reasoning performance among open-source models and strong results in specialized scientific tasks such as biomolecular interaction design and crystal structure generation. Two things are worth pulling out of that sentence. The first is the word jointly, which is the interesting one. Training twenty or more domains together is a claim that the domains reinforce each other, that a model that learns chemistry, materials science and biology in one mixture transfers between them. If that holds, the model generalises in a way a model trained on one domain at a time does not, and the payoff is data efficiency and generality rather than peak score in any single field. The second is the pairing of general reasoning with specialized generation. Biomolecular interaction design is a generative design task where a wrong output is expensive, and crystal structure generation is the same problem in a different domain, since a predicted crystal structure has to be physically plausible. The README also mentions, in the section on the 35B model, that it is the first open-source model with both crystal structure generation capability and strong general capabilities, which is the same trade-off stated from the other side: a specialist model can generate better, and a general model that can also generate is harder to build. None of this comes with numbers in the README, so the evaluation is in the reports and the leaderboards, not here. ## Long-horizon agents trained in sandboxes The third feature is the agentic one, and it is the part that most directly affects how you would use the model. By connecting multiple agent frameworks to large-scale sandboxed environments for black-box agentic reinforcement learning, the README says the model improves generalization and raises the capability ceiling for long-horizon tasks in both general and scientific domains. The 35B entry says the same thing more specifically, that it significantly improves agentic abilities over the previous generation, achieving strong results on multiple scientific agent benchmarks. The mechanism is worth understanding because it differs from ordinary reinforcement learning with a verifiable answer. In a sandbox, the agent runs against an environment that behaves like the real thing, the outcome is observed through the environment rather than through a reward function computed from a known answer, and the training signal comes from that interaction. That is black-box reinforcement learning, and it is what lets a model learn a long sequence of steps, use tools, and recover from its own mistakes. The phrase that multiple agent frameworks covers is the interesting constraint. Rather than teaching one harness, the training connects several, which is a way of avoiding a model that has learned to work only inside the shape of one particular scaffold. The cost is that the environments have to be good enough to be worth learning in, and building large-scale sandboxed environments for scientific work is itself a large project. For a user, the practical question is which agent framework the model has been trained against, and the README does not answer it. ## The 35B preview trades parameters for task coverage Intern-S2-Preview-35B is the most interesting entry in the README, because it makes an explicit argument about how to spend a training budget. It is described as an efficient 35B scientific multimodal foundation model, and beyond conventional parameter and data scaling it explores task scaling: increasing the difficulty, diversity and coverage of scientific tasks to push the model capabilities further. The concrete claim is that by extending professional scientific tasks into a full-chain training pipeline from pre-training to reinforcement learning, the 35B model achieves performance comparable to the trillion-scale Intern-S1-Pro on multiple core professional scientific tasks, while using only 35B parameters. It was continued pretrained from Qwen3.5, and it maintains strong general reasoning, multimodal understanding and agent capabilities. That is a claim about diminishing returns in parameter scaling, and it is the most falsifiable statement in the README: comparable on core professional scientific tasks, at a fraction of the parameters, with a different starting checkpoint. Whether it holds depends on which tasks count as core, which is exactly the kind of question the technical report has to answer. Two training details are described. During reinforcement learning the model adopts shared-weight MTP with KL loss to reduce the mismatch between training and inference behaviour, which the README says substantially improves the MTP accept rate and token generation speed, and it introduces CoT compression techniques to shorten responses while preserving strong reasoning capability, with improvements in both performance and efficiency. The first is about a speed-versus-training-fidelity trade-off, the second about the number of tokens spent thinking. The model also strengthens spatial modeling for small-molecule structures and introduces real-valued prediction modules, which is how it becomes, in the README's phrasing, the first open-source model with both crystal structure generation capability and strong general capabilities. ## S1-Pro in numbers, and what a reader should verify Intern-S1-Pro is documented in the third collapsed section, and it is the one model in the series with its parameter arithmetic spelled out. It is a trillion-scale mixture-of-experts multimodal scientific reasoning model, scaling to 1T total parameters with 512 experts, activating 8 experts per token, which works out to 22B activated parameters. The README says it delivers top-tier performance on advanced reasoning benchmarks and leading results across key AI4Science domains, chemistry, materials, life-science and earth among them, while maintaining strong general multimodal and text capabilities. The 1T total with 22B active shape is the standard trade in a mixture-of-experts model, and it has a direct operational meaning: the weights are large, the per-token compute is not, and the memory footprint is set by all 512 experts rather than by the 8 that are used. A team reading this should plan for a multi-node deployment or a served API, not for a workstation. What a reader should verify, then, is straightforward. The performance claims are the authors' own and the README carries no evaluation table, so the two technical reports it links, arXiv 2508.15763 for S1 and 2603.25040 for S1-Pro, are the documents to read. The model cards on Hugging Face and ModelScope are where the weights, the licence for each specific release and the evaluation details live, and the Apache-2.0 licence in this repository covers the README rather than the weights. The README also states that a chat endpoint is available, which is the cheapest way to form a view of the model before planning around it. The repository publishes no GitHub releases, and the last push was on 2026-09-13, so the README is the version of the announcement you are reading rather than a frozen one. ## Conclusion Use the Intern-S series when your problem is a scientific one that involves reading literature pages, designing molecules or generating crystal structures, because that is where the training effort described here is aimed. Do not treat the GitHub repository as an implementation you can build on, because it contains only READMEs, assets and docs, and the weights and any serving code are distributed elsewhere. Before you plan around these models, read the two technical reports the README links, arXiv 2508.15763 for S1 and 2603.25040 for S1-Pro, since the performance claims in the README are the authors' own, check the model cards on Hugging Face for the exact weights and licences of each release, and confirm the hardware you need, because a 397B model and a 35B model are not the same deployment decision. ## FAQ ### What is the Intern-S1 repository? It is an announcement and documentation hub for the Intern-S series of multimodal scientific models. The top-level listing contains only a README pair in English and Simplified Chinese, an assets directory, a docs directory and the Apache-2.0 licence, with the weights hosted on Hugging Face and ModelScope. ### What is the difference between Intern-S2-397B and Intern-S2-Preview-35B? The 397B model is the large multimodal foundation model scaled along pre-training, reinforcement-learning task coverage and agent environments, while the 35B model is the efficient one that explores task scaling instead of parameter scaling, was continued pretrained from Qwen3.5, and is reported as comparable to the trillion-scale S1-Pro on multiple core professional scientific tasks. ### What does Intern-S1-Pro scale to? Intern-S1-Pro is a trillion-scale mixture-of-experts multimodal scientific reasoning model with 1T total parameters and 512 experts, activating 8 experts per token, which is 22B activated parameters. The README links a technical report for it on arXiv as 2603.25040. ### How do I download the Intern-S1 or Intern-S2 models? The README links Hugging Face collections for both the Intern-S2 and Intern-S1 model families, a ModelScope collection, and an online chat endpoint. Nothing is downloaded from the GitHub repository itself, which contains no weights. ## Sources - [InternLM/Intern-S1 on GitHub](https://github.com/InternLM/Intern-S1) - [Issues](https://github.com/InternLM/Intern-S1/issues) - [License: Apache-2.0](https://github.com/InternLM/Intern-S1/blob/main/LICENSE) - [README](https://github.com/InternLM/Intern-S1/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/internlm-intern-s1