Confirmed in a primary source 13 Sept 2026 Models

A Chinese government laboratory released a 397-billion-parameter model on Sunday and attached no conditions to its use.

The Shanghai Artificial Intelligence Laboratory published Intern-S2-397B on Hugging Face at 03:20 UTC on Sunday: 397 billion parameters, a mixture-of-experts design that reads images as well as text, aimed at scientific reasoning and long-running agent tasks, under the Apache 2.0 licence. Apache 2.0 is the permissive end of open source, which means anyone can download the weights, modify them and sell products built on them, with no approval, no usage restrictions and no geography clause.

The caveat is that a release is not an adoption: the repository had 23 likes and zero recorded downloads when we checked, the laboratory has published no independent benchmark results we could verify, and a model of this size needs serious hardware before anyone can judge whether it is any good.

Until now

Open-weights releases at frontier scale were treated as a Chinese company strategy, run by DeepSeek, Alibaba, Moonshot and MiniMax competing for developers.

Now

Read frontier-scale open weights as state research output as well as company strategy, and expect the most permissive licences in the field to keep coming from institutes rather than from the labs arguing about how fast to go.

It went up at twenty past three in the morning, Sunday, UTC, with no announcement anyone noticed while the industry was reading an essay about slowing down. Intern-S2-397B is a 397-billion-parameter mixture-of-experts model that takes images and text and returns text, published by the Shanghai Artificial Intelligence Laboratory under Apache 2.0. The laboratory says it was pre-trained visually on raw pages of scientific papers and tuned with reinforcement learning across more than twenty scientific fields, and that it is meant for scientific reasoning and agents that run for a long time without supervision. It replaces a preview version from July. The licence is the part that matters most and the part least likely to be reported: Apache 2.0 carries no acceptable-use policy, no field-of-use restriction and no clause about where the user is.

The honest limits are large. Zero downloads and twenty-three likes at the time we checked is not adoption, it is a file on a server. We have no independent evaluation of the model, only the laboratory's own framing, and a model this size takes a cluster before anyone can form a view. What can be said without stretching is the asymmetry of the weekend. The most-discussed events of September 12 and 13 were four American chief executives agreeing that capability should advance more slowly and promising outside inspectors. The largest thing actually published in those two days was a Chinese state institute putting 397 billion parameters on the open internet under a licence that lets any company anywhere build a business on it. Read frontier-scale open weights as state research output as well as company strategy, and expect the most permissive licences in the field to keep coming from institutes rather than from the labs arguing about how fast to go.

How it got here

Jul 16
The laboratory publishes Intern-S2-Preview-397B, the preview version this release supersedes.
Sep 11
The FP8 variant of Intern-S2-397B is created on Hugging Face, ahead of the flagship.
Sep 12
Dario Amodei publishes his pacing essay; Altman, Musk and Hassabis agree the same day. Carded in Edition 020.
Sep 13
internlm/Intern-S2-397B is created at 03:20:35 UTC under Apache 2.0, 397 billion parameters, multimodal. Xi Jinping proposes a BRICS open-source AI initiative at the New Delhi summit the same day.
Who has to react
Every startup that cannot afford frontier API bills and now has a 397-billion-parameter multimodal model for free; the American labs whose pricing assumes a capability gap at the top; the six Chinese firms named in the NSA, CISA and FBI distillation advisory, because a permissive Chinese release complicates the argument about who copies whom; Hugging Face, which hosts it and is being bought by Nvidia; export-control officials in Washington, for whom weights published under Apache 2.0 cross every border at once; and Chinese policymakers, whose leader proposed an open-source AI initiative to BRICS the same day.

The evidence

ObservedThe repository internlm/Intern-S2-397B was created on Hugging Face at 2026-09-13T03:20:35Z and last modified at 2026-09-13T03:49:44Z; we read the Hugging Face API record. Licence: apache-2.0.

The whole observation, as recorded

Pipeline: image-text-to-text. Architecture: Qwen3_5MoeForConditionalGeneration, a mixture-of-experts model, tagged qwen3_5_moe and safetensors. Parameter count in the model name: 397B. Downloads at the time of checking: 0. Likes: 23. The laboratory's stated positioning is scientific reasoning and long-horizon agents, with visual pre-training run directly on raw pages of scientific literature and reinforcement learning across more than twenty scientific domains; its stated evaluation configuration is a 256K maximum inference length for text reasoning and 64K for multimodal. An FP8 variant, internlm/Intern-S2-397B-FP8, was created on September 11 or 12; the full-precision flagship is the September 13 event. It supersedes Intern-S2-Preview-397B, published on July 16, 2026. The Shanghai Artificial Intelligence Laboratory is a state-backed research institute, and its InternLM model family has been released openly since 2023. For scale: this is the largest open-weights release of the window anywhere in the world, and nothing comparable came out of any American or European lab on September 13 or 14.

Established

The Shanghai Artificial Intelligence Laboratory published Intern-S2-397B, a 397-billion-parameter multimodal mixture-of-experts model, on Hugging Face at 03:20:35 UTC on September 13, 2026.
It is licensed under Apache 2.0, which permits commercial use, modification and redistribution with no field-of-use or geographic restriction.
It supersedes Intern-S2-Preview-397B, published July 16, 2026, and an FP8 variant created two days earlier.
No comparable open-weights release came from any American or European laboratory on September 13 or 14, 2026.

Not established

That the model is good. There is no independent evaluation, no third-party benchmark and no leaderboard placement we could verify.
That anyone is using it. The repository recorded zero downloads and 23 likes when we checked.
That it reflects Chinese state policy. A state-backed institute publishing research is not the same as a government directive, and none has been published.
That it changes the balance of capability. Parameter count is not capability, and the laboratory's own claims about scientific reasoning are untested outside its walls.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityHard

PersistenceStructural

Cost is high because training a 397-billion-parameter multimodal model costs real compute and real staff, and giving the weights away forecloses ever charging for access to them. Commitment is high because publishing weights is a one-way act; there is no version of this that is a trial. Reversibility is hard in the strongest sense available: the file can be deleted from Hugging Face and it cannot be deleted from the machines that pulled it, and the Apache 2.0 grant to anyone who already has it is irrevocable. Persistence is structural for the same reason, and it is the one dial here that deserves a full mark. Everything about this signal is weak except the fact that it cannot be undone, which is precisely why it belongs in the deck on a quiet Monday.

What to watch

strengthens By March 31, 2027, a second Chinese state-backed institute publishes an open-weights model of 200 billion parameters or more under a permissive licence, or a Chinese ministry names open-weights release as a policy instrument in a published measure.
weakens Intern-S2-397B records fewer than 10,000 downloads on Hugging Face by December 31, 2026, or the repository is withdrawn or relicensed.

The words

Open weights: The trained model file itself is published, not just access to it through a website. Anyone who downloads it can run it on their own machines with no connection to the publisher.
Apache 2.0: A permissive software licence. It allows commercial use, modification and redistribution with almost no conditions, and unlike many model licences it puts no limits on who may use it or for what.
Mixture of experts: A design that splits a model into many specialised sub-networks and uses only a few for each input. It makes very large models cheaper to run than their total parameter count suggests.
Parameters: The numbers a model learns during training. 397 billion of them is frontier scale, and puts this in the same size class as the largest openly published models.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.