Mistral gave away a model the size of a trillion.

The French company published the weights for Large 4, nicknamed Le Chonk. Anyone can download it. That is not the same thing as open source, and the gap is where the interesting part lives.

On 29 October, Mistral published the weights for Mistral Large 4, about three weeks after letting people use it through an API. The company's own nickname for it is Le Chonk, which tells you roughly how big it is.

You can download it. That is the news. What is worth a few minutes is what you are downloading, and what you are not.

What a weight is

A model is a very large pile of numbers. Training is the process of nudging those numbers until the thing behaves. When it is finished, the numbers are the model — there is nothing else in there.

Those numbers are the weights. Publishing them means handing over the finished article.

Open weights is not open source

This is the distinction the phrase is doing a lot of work to blur.

Open source, properly, means you get what you would need to rebuild the thing: the code and the data. Open weights means you get the result. The training data and the training code stay with the company.

It is the difference between being given a cake and being given the recipe. The cake is genuinely useful. You still cannot make another one.

Mistral released this under its own licence, the Mistral Ecosystem Open-Weight License, which at least has the honesty to say so in the name.

Who this is actually for

Anyone who cannot send their data to someone else's server. Governments, hospitals, banks, defence. Mistral has leaned on that argument since it started, and “sovereign control” is the phrase its own announcement reaches for.

What it does not do is make the model cheap. Downloading a trillion-parameter model and running it are different problems, and the second one still needs a room full of hardware. Mistral points self-hosters at vLLM or SGLang.

The numbers, and a caveat

Mistral reports 41% on Terminal-Bench 4.0, which it says is thirteen points above the preview build, and credits three further weeks of reinforcement learning.

Treat that as the company's own figure until someone else reproduces it. Reported specifications also differ between outlets — the active parameter count is given variously as 49 or 52 billion — so the precise numbers are softer than they look.

The idea to take away

A mixture-of-experts model is mostly asleep. About a trillion parameters exist; only around fifty billion of them are consulted for any given word. That is how a model this size can be run at all.

And when a company gives one away, read the licence before the benchmark. The benchmark is a claim about the model. The licence is a claim about you.

■