Mistral Large 4: Europe plants its flag in AI

Not the fastest model in the world. But trained from scratch in Europe, and that is the whole point.

French lab Mistral has released the open weights of Large 4. It is a big one, more than a trillion parameters, built as a mixture of experts so only about 52 billion of them work on any given request. It handles a million tokens of context. And it was trained from scratch in European data centres.

Let us be honest about the specs. A million-token window and mixture of experts are standard kit these days; Grok, Gemini and DeepSeek have them too. On raw performance it does not catch the top American models. And "open" comes with a catch: you can download and use it for free, but bigger commercial uses have licence limits, so it is not fully open in the Apache sense.

So if it is not the best and not even fully open, why should anyone care?

Because the story is not about benchmarks. It is about a flag.

For years the AI race has looked like a two-player game: the United States and China. Everyone else was supposed to pick a side and rent their intelligence from someone else's servers. Mistral is saying no. Europe can train its own frontier-scale model, on its own hardware, on its own soil.

That matters more than a couple of points on a leaderboard. If your public services, your banks, your hospitals all run on models controlled from abroad, you are dependent, full stop. A home-grown model, even one a step behind, is insurance. It is leverage. It is the difference between being a customer and being a player.

So no, Large 4 will not top the charts. It does not need to. It is Europe defending its honour, and showing it has not been left behind.