Mistral Large 4 Is Out, but Le Chonk's Weights Are Not

Mistral released Mistral Large 4 on Tuesday, a 1-trillion-parameter multimodal model — but the weights are not downloadable yet, and the company's own accounts of when they will be differ.

The French lab nicknamed the model Le Chonk, and is pitching it at buyers who do not want to depend on an American or a Chinese lab. TechCrunch reports the release came "following what French president Macron described as 'a third way in AI.'" For anyone evaluating it this week, the practical question is narrower: what can you actually run, and when.

The weights are not out yet

This is where the reporting splits, and it matters if you were planning to self-host.

TechCrunch is explicit that the model "is not an open-weight model yet," adding that "for the time being, it can only be accessed via a public guardrail endpoint, but Mistral plans to make its weights available in just three weeks, after safety testing is complete." Mistral VP Science Pierre Stock framed the delay as deliberate: "In the meantime, we'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks."

WIRED describes the same release differently, reporting that the model "can be used and customized by anyone" and is "currently available in preview, with a final version to follow by the end of the month." Vercel, announcing gateway support the same day, calls it simply "the open-weight, natively multimodal model."

Read together, the one claim every source supports is that a final, downloadable release is weeks out, not days. If your plan depends on holding the weights — air-gapped deployment, a regulated workload, a fine-tune — treat late October as the earliest honest date and build nothing on the preview that you cannot rebuild.

4,000 GPUs, and a from-scratch claim

Mistral's sharpest technical claim is about cost, not capability. According to TechCrunch, Mistral Large 4 "was trained entirely on Mistral's compute; using only 4,000 Nvidia GPUs," which Stock characterised as "two to three times less than our Chinese competitors, and significantly less than the closed source competitors."

That figure is the company's own and has not been independently verified. It is worth separating from the second claim stacked on top of it: WIRED notes that where "Chinese labs have been accused by the US government of abusing distillation — the training of a smaller model on the outputs of a larger one — to close the performance gap with OpenAI and Anthropic, Mistral claims to have trained its model from scratch."

Both statements are Mistral's position, and both are the kind of thing the eventual weight release will let outsiders probe. Until then they are marketing with a plausible shape.

On performance, Mistral is not yet showing its work. TechCrunch reports that "benchmark results [are] still pending," while Mistral "hopes ML4 will be best in class among open-weight models, especially outside of China." WIRED carries the stronger version of the pitch: Mistral presents the model "as by far the most capable open-weight model developed outside of China, and 'very, very close' to some proprietary models." Nothing public yet tests either framing.

Tuned for niches, not leaderboards

The model's optimisation targets are unusually specific, and they line up with who pays Mistral's bills.

WIRED reports it is "optimized specifically for coding and cyberdefense, as well as tasks particular to manufacturing, finance, electrical engineering, and other niches." Cofounder and chief scientist Guillaume Lample put the strategy plainly: "There are a lot of areas where the other labs will not focus that much. There are so many domains in which you can improve models."

TechCrunch adds chip design to that list and connects it directly to the cap table: optimised use cases "include cybersecurity and finance, but also chip design, which is core to two of Mistral's main backers — Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation (about $24.39 billion)."

That is a useful filter. A model tuned around cyberdefense and electrical engineering is not aiming to beat a frontier generalist on a chat leaderboard, and judging it on one would miss the point.

How to call it today

You do not have to wait for weights to run an evaluation. Vercel shipped Mistral Large 4 on its AI Gateway the same day, and says you "can access the model through AI Gateway with one API key, without creating a separate Mistral account."

Per Vercel's changelog, the model id is mistral/mistral-large-4, usable "with the AI SDK, OpenAI-compatible Chat Completions API, Responses API, or Anthropic Messages API," and selectable "in a coding agent connected to AI Gateway." If you already route traffic through an OpenAI-compatible endpoint, that is a model-string change, not an integration.

The sensible test this week is a narrow one: take a workload in Mistral's stated strike zone — a code task, a log-triage or cyberdefense prompt — and run it head to head against whatever you use now. Do not generalise the result to the final weights, which is a different build.

The argument underneath the launch

Mistral's commercial case is less about benchmarks than about continuity of supply.

WIRED notes that "in June, the Trump administration placed temporary restrictions on the distribution of models from OpenAI and Anthropic, citing concerns they could be abused to launch sophisticated cyberattacks." Lample's framing follows from that: "Sometimes, people like to [make a big deal] over the US, versus Europe, versus China. But what really matters is to own the model — even for US companies. If you use a closed model, there is no guarantee it will still be there tomorrow."

There is a cost argument stacked behind it. WIRED observes that it is "already considerably cheaper for businesses to run open-weight models, which cost only as much as the compute they consume."

The company has the balance sheet to press the point. Per WIRED, Mistral "raised a $3.3 billion funding round at a $24 billion valuation" in September, "the largest ever raise by a European tech company," with earnings that "have reportedly increased 20-fold in the last year or so."

What to watch

• Late October. TechCrunch's three-week timeline and WIRED's end-of-month final version both land there. If the weights slip past it, the safety-review framing becomes the story.
• The first third-party benchmarks. Mistral has published none. The "best open-weight model outside China" claim is untested until someone outside the company runs it.
• What the licence actually permits. Stock's wording — weights usable "to defend, but not to [perform] malicious attacks" — implies terms that constrain use. Read them before you plan a deployment around them.
• Whether the 4,000-GPU figure survives contact. It is the most quotable number in the launch and the least verifiable one today.