On Tuesday morning Mistral put a public preview of Mistral Large 4 behind an API: one trillion parameters, 49 billion of them active, natively multimodal, and — because this is a French company — officially nicknamed le Chonk. The weights come down by the end of the month, October 27 according to The Next Web. It’s the first model out of the €3 billion round Mistral closed in September.
That last detail is the one to hold onto. A €3 billion company does not release frontier weights out of charity. It releases them because the release calendar in Brussels now matters as much as the one in San Francisco.
A Release Calendar Written in Brussels
The pre-written take is that the AI Act’s general-purpose model obligations began biting on August 2, 2026, and that Europe’s champion is about to be strangled by its own regulators. Every trade association in Brussels has that op-ed in a drawer.
Sidestep it. Ask who actually carries the burden under the regime that started August 2. Not the company publishing a blog post. The company that puts the model into production.
The Commission’s carve-out for free and open-source releases is real: providers who publish parameters, architecture information, and usage documentation get relief from a chunk of the Article 53 paperwork. The carve-out also comes with an exception, stated plainly in the Commission’s guidance — it does not apply to models presenting systemic risk. Systemic risk is presumed above a training-compute threshold of 10^25 floating-point operations, with the Commission free to designate on other grounds. Mistral has published a parameter count, a nickname, and a benchmark chart. It has not published its training compute. Nobody outside the building knows where ML4 sits relative to that line.
The Carve-Out Was Drafted for Somebody Else
Here’s the part both sides keep stepping over. The open-source exemption was written with a mental image: a lab, a repository, a grad student, maybe Meta feeling generous. It was not written for a company with a €3 billion war chest and a 125-enterprise customer list that includes HSBC, ASML, and Airbus.
That is not a loophole Mistral invented. It’s a mismatch Brussels walked into. The mismatch has a bill attached, and the bill doesn’t arrive at Mistral’s address.
Under the August regime, the duties that follow a general-purpose model — transparency toward downstream users, copyright compliance, documentation, incident reporting, cybersecurity of the model and the infrastructure it runs on — travel with the artifact. When the artifact is a downloadable file on your own servers, the practical work of “documenting serious incidents” and “ensuring adequate cybersecurity protection” belongs to whoever deployed it. That’s a procurement decision with a legal department attached, and most procurement departments haven’t been told.
“We can read the license,” a procurement lawyer for a European manufacturer told me, on a call from a courthouse hallway in Düsseldorf. “We still can’t read the model. At least with a closed vendor I have somebody to sue.”
What Open Weights Buy, and What They Cost
Mistral’s strategy is not naive. Open weights buy distribution, standards influence, hiring gravity, and a hedge against the morning an American API decides European customers are a sanctions problem. It is the smartest available play for a company that cannot out-spend OpenAI on closed-model training runs and whose next tranche of compute doesn’t come online until the first half of 2027.
But cheering it as a blow for freedom misses what’s being transferred. Hacker News greeted the launch with 1,629 points and 978 comments, most of them variations on “open weights are a public good.” At a trillion parameters, with those customers, open weights look more like a liability transfer — regulatory obligation handed downstream along with the tensor file, wrapped in the language of liberation.
The uncomfortable part for the right is this: the market’s answer to a Brussels rulebook was not to fight it. It was to privatize it. The duties simply moved from a company with a legal team to clusters inside companies that also have one — and a long tail that doesn’t.
So the interesting question isn’t whether Europe can build a frontier model. It plainly can, and it just did. The question is who signs the incident report at 2 a.m. when those weights are already copied onto somebody’s unpatched box — and why we spent three years arguing about whether open source is a philosophy when it was, all along, a compliance instrument.
Sources
- Mistral AI launches Large 4 preview ahead of open-weight release
- Introducing Mistral Large 4 | Mistral
- Mistral releases Large 4, a 1 trillion-parameter open-weight AI model
- EU AI Act GPAI: August 2026 Enforcement Decoded
- EU AI Act GPAI enforcement begins August 2, 2026 · AlCybr
- EU AI Act GPAI enforcement begins August 2026 | ComplianceHub.Wiki