Mistral's open Large 4 is Europe's best sovereignty argument, with one catch

Can you be sovereign in AI if your best model lives on someone else's server and can be switched off? Europe has spent years giving a vague answer. On October 29, Mistral gave a concrete one: you download it.

Mistral has published the weights of Large 4, its flagship, on Hugging Face, along with a quantized version. The model has roughly a trillion parameters, of which 52 billion are active per generated token. According to Frandroid, it scores 41% on Terminal-Bench 4.0, up from 28% for the preview three weeks earlier, and ties GLM-5.3, the best Chinese open model on Mistral's chart, on Artificial Analysis's agentic coding index. Those figures come from charts Mistral showed the press, so treat them as the vendor's own until independent tests confirm them.

The weights matter more than the leaderboard

Leaderboards expire. Google just put out Gemini 4 Carbon, which tops Artificial Analysis's Intelligence Index with 59 points, and for now it is reserved for a handful of cybersecurity specialists, with no date for anyone else. That is the point. A closed model is a service. Whoever runs it decides who gets access, when and at what price.

A downloaded checkpoint is different. Corentin Petit, Mistral's VP in charge of solutions, pitches an on-site cyberdefense tool that, unlike closed models, nobody can cut off. We think that is the strongest argument for European sovereignty so far. It beats "we built our own chatbot" and it beats a point of benchmark share. Sovereignty means your hospital, bank or ministry keeps working when a foreign vendor changes its terms, its safety policy or its mind.

The cost side helps the case too. Mistral says Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in its European data centers, and CEO Arthur Mensch contrasts its roughly $6 billion raised in three years with the hundreds of billions other labs have raised. Leaving the exact figures aside, the message is that a European lab can reach the frontier of open models without American-scale budgets.

"Open" has a fine print, and a power bill

Now the catch we flagged in the headline. The license is not Apache 2.0, like the small Ministral models. It is a modified MIT: you can use, modify and commercialize the model, except if you resell access to it as a service, in which case you need a commercial license from Mistral. For nearly every company building products on it, that is generous. For European cloud providers that wanted to become the sovereign host of this model, it is a real limit, and we think it should be debated more loudly. A sovereignty strategy that routes every hosting business back through one company's sales team is only half open.

Then there is hardware. Even in 4-bit form, the model weighs around 610 GB and, according to Mistral, needs four Nvidia B200 or B300 GPUs. That is data-center gear, not something a startup keeps under a desk. The "anyone can download it" claim is true and also a bit theatrical. In practice, the people who benefit first are probably large enterprises, governments and well-funded hosts.

Mistral promises a family of smaller models within the month, one of which will certainly run on a laptop, according to Pierre Stock, its VP in charge of science. But it gave no size, no firm date beyond that and no test results, and it did not say what "laptop" means. A model that needs 128 GB of unified memory is not the same promise as one that fits a 16 GB machine. We would hold our applause until we see the memory requirement and the license.

There is also an irony we cannot skip. The model was trained on Nvidia hardware, and Hugging Face, where the weights live, is a platform Nvidia announced in September that it would acquire, as Frandroid notes. European sovereignty at the model layer still sits on an American stack underneath. That does not cancel the achievement, but it is a reminder that open weights are a necessary condition, not a sufficient one.

The safety question is the next test

The least comfortable part is cybersecurity. Large 4 scores 79% on CyberGym-E2E, a test that asks the model to reproduce a real vulnerability and then fix it, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task. Mistral withholds a less restricted cyber variant for contracted partners and says the public version refuses 96.6% of malicious requests. It also admits it saw potential for helping real attacks in its API traffic and is tightening safeguards, but that only applies to models it serves. Downloaded weights do not update. Pierre Stock did not detail which safeguards ship in them.

Our view: that is the honest price of open weights, and the right response is not to close them but to be candid about it and let independent researchers test the published checkpoint.

Verdict

Large 4 is the clearest proof yet that Europe can field a frontier-adjacent open model. If you run infrastructure in Europe, evaluate it now, but read the Model-as-a-Service clause first. If you are an individual, wait for the small family and judge it on memory footprint and license. And if you are a policymaker, note the lesson: sovereignty is something you can download, but only if the license, the hardware and the supply chain let you.