Its creators have nicknamed it "Le Chonk". On the internet, the word affectionately refers to a cat that has been a little too well fed.
Here, it refers to the biggest artificial intelligence model ever built by Mistral, the French start-up that has become the symbol of European AI. Its official name is Mistral Large 4, and it was unveiled on Tuesday.
What Large 4 is
It is a model of around 1 trillion parameters, the internal settings that determine its answers. But it only uses around 49 billion of them at a time: depending on the question, only part of the model gets to work, a bit like a large hospital where only the relevant specialists are consulted. We explained this principle in our article on the "mixture of experts" architecture.
It understands text and images natively, and it was trained on content in more than 160 languages. Mistral built it from scratch, on 3,800 Nvidia chips installed in its own data centres in Europe, and it is also offered from Europe.
For now, this is a public test version, accessible to developers. Mistral promises to release its weights, that is, the model itself, downloadable by anyone, by the end of the month, after a series of safety tests carried out with cybersecurity specialists, partners and public authorities. Its price has not yet been announced.
What Mistral claims
The company presents Large 4 as the best open model to come out of the United States or Europe, across all tests. It says it is at the cutting edge for cyberdefence, industry and finance, and even superior to the best closed models at pinpointing elements in an image.
As always, these claims come from the company itself. An independent ranking gives a more measured picture: according to the Artificial Analysis index, Large 4 scores 38 points, against 9 for its predecessor Large 3. The leap is spectacular. But Claude Opus 5.5, one of the best American models, scores 58.
Mistral compares itself to American and European open models. That is not a detail: today, the most powerful open models often come from China, such as those in the GLM, Qwen or Kimi families. Being the best open model outside China is a real achievement, and a strong argument for companies that do not want to depend on a Chinese model. But it is also a reminder that the race for open models is largely being played out elsewhere.
Why it matters for Europe
Large 4 is not just one more model. It is a model designed, trained and hosted in Europe, end to end.
For a European company or public authority, this is not just a matter of pride. It is a matter of law: data processed by an American company can, in some cases, be demanded by American courts, even if it is stored in Europe. We explained this in our article on the Cloud Act. A European model, hosted by a European company, avoids this problem.
And if it is released openly as promised, anyone will be able to run it on their own machines, without depending on anyone.
What we take away
Mistral Large 4 is not the best model in the world. It does not need to be. As we wrote in September, mattering does not require being first.
What Large 4 brings is a choice that did not exist: a very good model, soon to be open, made and hosted in Europe. For many European companies, that is exactly what was missing. See you at the end of October, to check that the promised weights really turn up.