Can One Fat Cat Put Europe Back in the AI Race?

Europe's biggest AI model landed this week with a joke for a name, a trillion numbers to its back, and a promise to hand the whole thing over. The question is whether it can back the brag.

Mistral, the French lab, calls it Mistral Large 4. Everyone else calls it Le Chonk, the internet's word for a very fat cat. It is one trillion parameters large. If that means nothing to you, picture a machine with a trillion tiny dials that it learned to set by reading a small library. Most models that size stay locked on a company's servers. Mistral says it will give this one away.

An open model is one you can download, run on your own machines, and change to suit your work. A closed model is one you can only rent, and the company can switch it off. Mistral's bet is that enough businesses want the first kind.

The numbers, and the small print

The headlines are strong. Mistral reports 61.7 per cent on a hard coding test, a top-five finish in an independent cyber ranking, and 82 per cent on a security test where the model has to find a real flaw in software and then fix it. It passes 15 per cent of a legal benchmark. On finance, 67 per cent.

Now the small print. These figures are early and reported by Mistral itself. Rival scores in its own charts do not always match what those rivals have published. On the live coding leaderboard, some open rivals score higher, and several closed models sit around 74 per cent. Le Chonk has not yet appeared in any independent test.

A trillion parameters is a headline. Forty-nine billion awake is the trick.

Made in Europe, on purpose

Le Chonk was trained from scratch on about 4,000 of Nvidia's best chips, in Mistral's own data centres in Europe, over two months. It was fed more than 160 languages, including every official language of the European Union.

That is the point. In the past year, Washington has limited how the best American models can be shared, and asked US labs to hold new models back even from Britain's safety testers. For any business whose cyber-defences rest on an American service, the lesson is uncomfortable: the model you depend on might not be there tomorrow. Mistral's answer is a model you own.

The joke that came true

In June, a made-up Mistral model called Le Chaton Fat went viral, complete with fake charts and silly specs. Some versions promised 1,000 meows per second. The chief executive joined in. By July, people genuinely expected a giant model. This week, they got one. The company is not above the joke, and it is happy to prove the jokers right.

The honest counterpoint

Mistral has a fair defence. Open models are a strategy, not charity: China proved they win users, and a European lab matching them is a reasonable answer. The numbers may well hold once the files are out. What would change my mind: if outside testers reproduce Mistral's scores, my caution softens. If they cannot, the boast was the product.

What it means if you build on AI

I build local and on-prem systems for a living, so I am not neutral. The lesson is not to cheer for a flag. It is to keep the layers you depend on within reach. A trillion-parameter model you can run yourself is a real option, not a slogan, and options are the thing worth having.

If you are weighing what to own in your own AI stack, email brandon@kreostudio.co.uk.