News · 2026-09-13

InternLM releases Intern-S2, a 397-billion-parameter open science model that is a 406 GB download even in FP8

The InternLM team has released Intern-S2-397B, an open-weights multimodal model built for scientific reasoning and long-running agent tasks, under the permissive Apache 2.0 licence. It is a very large mixture-of-experts model: the FP8 version, the one most people would deploy, is a 406.3 GB download, and InternLM recommends a server node with eight top-end data-center GPUs to run it.

Key facts

• Size: 397 billion parameters; a 406.3 GB download in FP8 and 806.9 GB in bf16, according to Hugging Face's file listings.
• Hardware: InternLM's deployment guide says: “We recommend deploying the Intern-S2-397B model on H100 (x8) or H200 (x8) nodes,” that is, eight 80 GB or eight 141 GB GPUs.
• When: the FP8 repository appeared on 11 September 2026 and the bf16 repository on 13 September, following a preview release in July.
• Primary sources: the model card, the FP8 weights and the InternLM GitHub repository.

A science model, not a chatbot

Most open-weights releases compete on general chat and coding. Intern-S2 targets a narrower and harder audience. InternLM describes it as “our most capable multimodal foundation model for scientific intelligence and long-horizon agents.”

Its distinguishing training choice is how it reads research. The model card says Intern-S2 “learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships in a shared representation space without intermediate parsing.” In plain terms, instead of first converting a paper's PDF into clean text and losing the equations, diagrams and tables in the process, the model looks at pages the way a scientist does, figures and formulas included.

After that, InternLM applied reinforcement learning across “more than 20 domains,” with examples such as “biomolecular interaction design and material structure generation.” The inference engine LMDeploy also supports time-series inputs such as earthquake detection signals.

How it is built

Intern-S2 is a mixture-of-experts model. Its configuration lists 512 experts, with 10 active for each token, across 60 layers, and a native context window of 262,144 tokens. The analogy is a large hospital: hundreds of specialists are on staff, but any given patient sees only a handful. Only a small share of the model's parameters do work for each word, which keeps generation faster than a dense model of the same size, but every expert still has to sit in memory, which is why the hardware bill stays high.

That explains the two downloads. The bf16 version stores each number in 16 bits; the FP8 version uses 8-bit numbers and roughly halves the size, a form of quantization. For readers wondering why a single model arrives as hundreds of files, our lesson on model file formats explains the sharding.

The model can be served with LMDeploy (version 0.14.0 or later), vLLM (0.22.1 or later) or SGLang (0.5.13 or later). The deployment guide's vLLM example splits the model across eight GPUs and stretches the context to about a million tokens using YaRN, a standard technique for extending how far a model's position encoding reaches, while its LMDeploy example uses a 512,000-token context. Those long contexts come on top of the weights: the memory that stores each conversation, explained in our lesson on the KV cache, grows with every token.

Why it matters

Open science models let research groups run, inspect and fine-tune a capable system on sensitive or unpublished data without sending it to a commercial API. Intern-S2 joins a small field that includes Arcee's Genesis Science 1, and the Apache 2.0 licence allows commercial use with few restrictions.

It is also a marker of how open frontier-scale models are distributed now: not as a laptop download but as infrastructure for institutions with a GPU cluster.

The caveat

The benchmark results on the model card are published only as an image, so there are no text figures to check, and no independent evaluations had appeared at release, when download counts still stood at zero. The card's text closely follows that of Intern-S2-Preview-397B, released on 16 July, and InternLM has not published a separate announcement explaining what changed between preview and release. The eight-GPU recommendation also means very few individuals can run it at all.

Primary source, verified: read the paper →

Ground Truth. (2026, September 13). InternLM releases Intern-S2, a 397-billion-parameter open science model that is a 406 GB download even in FP8. Ground Truth. https://groundtruth.day/news/intern-s2-397b-ships-as-a-406-gb-open-weights-science-model.html

@misc{groundtruth:intern-s2-397b-ships-as-a-406-gb-open-weights-science-model,
  title  = {InternLM releases Intern-S2, a 397-billion-parameter open science model that is a 406 GB download even in FP8},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/intern-s2-397b-ships-as-a-406-gb-open-weights-science-model.html}
}

Topics: open-weights · model-release · science · mixture-of-experts · multimodal · apache-2.0