Intern-S1

Open-weight language models have closed much of the gap with closed frontier systems on general reasoning, but scientific work has lagged: the professional tasks that matter to a working chemist or biologist still tend to be served by narrow expert models, and general-purpose assistants handle molecular formulas, protein sequences, and instrument readouts as ordinary text. Intern-S1 is Shanghai AI Laboratory's answer — what the authors call a "specialized generalist", a multimodal model that keeps broad reasoning ability while treating scientific data formats as first-class inputs.

The design choice that distinguishes Intern-S1 from a general chat model is its dynamic tokenizer, which parses molecular formulas, protein sequences, and time-series signals natively rather than shredding them into ordinary subword tokens. Combined with a vision encoder for figures, spectra, and micrographs, this lets one checkpoint move between reading a paper figure, interpreting a chemical structure, and reasoning over a sequence. Roughly half of its continued-pretraining corpus is scientific.

Intern-S1 belongs to the same family of broad scientific models as Galactica, NatureLM, the DARWIN Series, and GPT-Rosalind, and it shares their central caveat: this is a multi-science system whose strongest published results are in chemistry and materials science, not a biology specialist. Its biological relevance is real but partial — protein-sequence understanding and molecular reasoning sit inside a much wider remit that also spans physics, earth science, and general reasoning.

#Key Features

• Native scientific tokenization: A dynamic tokenizer encodes molecular formulas and protein sequences as their own token streams, so chemical and biological strings are not fragmented by a tokenizer tuned for prose.
• Multimodal scientific input: A 6B InternViT vision encoder handles figures, chemical structure diagrams, and microscopy-style images alongside text in a single context.
• Sparse Mixture-of-Experts efficiency: Only 28B of the 241B total parameters activate per token, keeping inference cost far below what a dense model of the same capacity would demand.
• Reinforcement learning across a thousand tasks: Post-training uses offline then online RL in the InternBootCamp environment, with a Mixture-of-Rewards scheme that trains on more than 1000 heterogeneous tasks simultaneously.
• Permissive open weights: The full checkpoint is published on HuggingFace under Apache-2.0, alongside quantized variants and a smaller Intern-S1-mini for teams without large-scale serving infrastructure.

#Technical Details

Intern-S1 pairs a 235B-parameter Qwen3 Mixture-of-Experts language backbone with a 6B InternViT vision encoder, for 241B total parameters and 28B activated per token. It is continually pretrained on 5 trillion multimodal tokens, of which more than 2.5 trillion come from scientific domains. Post-training applies offline followed by online reinforcement learning, using the Mixture-of-Rewards approach to reconcile reward signals across more than 1000 tasks in one training run.

On the science benchmarks reported in the model card, Intern-S1 scores 83.4 on ChemBench, 75.0 on MatBench, 65.7 on MSEarthMCQ, 63.9 on MicroVQA, 63.1 on ProteinLMBench, 51.0 on SmolInstruct, and 44.3 on SFE. The paper reports that it exceeds closed-source state-of-the-art systems on molecular synthesis planning, reaction condition prediction, and thermodynamic stability prediction for crystals, while remaining competitive with other open models on general reasoning. The smaller Intern-S1-mini pairs a dense 8B Qwen3 backbone with a 0.3B InternViT encoder under the same license.

#Applications

For biological research the practical uses are literature- and reasoning-shaped rather than predictive: interpreting figures and structures from papers, reasoning over protein sequences and their annotations, planning synthetic routes to a compound of interest, and acting as an agentic assistant that stitches those steps together. Because the weights are Apache-2.0, groups can fine-tune the model on in-house scientific corpora, and the mini variant makes local experimentation feasible on modest hardware; serving the full 241B model, however, requires substantial GPU capacity, so most users will reach it through hosted inference.

#Impact

Intern-S1 shows that a fully open-weight model can compete with closed systems on professional scientific tasks rather than only on general benchmarks, and its Apache-2.0 license makes it a usable base for domain fine-tuning in a way that API-only frontier models are not. The limits are equally plain: the reported gains concentrate in chemistry and materials science, the biological evaluation rests largely on a single sequence-understanding benchmark, and the composition of the 2.5 trillion scientific tokens is not broken out by discipline. No training-data release accompanies the weights, and no data card documents the corpus.