Mistral Large 4 'Le Chonk': A Trillion Open-Weight Parameters, Made in Europe

A trillion parameters, downloadable for free, made in Europe. Mistral picked an interesting week to remind everyone that open-weight AI is very much alive.

On October 29, the French lab released Mistral Large 4, cheerfully nicknamed "Le Chonk" - and the name is earned. It's a mixture-of-experts model with around 1,000 billion total parameters, of which about 52 billion are active at any time, with a context window of one million tokens. And the weights are right there on Hugging Face.

Why this one matters

Big open releases have been getting rarer at the frontier. Most labs have retreated behind APIs, keeping their best models locked up and metered. Mistral just dropped its flagship into the open with day-zero support for vLLM and llama.cpp, which means people were running it on their own hardware within hours.

Mistral claims it's the best open-weight model for coding, using 47% fewer tokens than GLM-5.3 on equivalent tasks. Vendor benchmarks are vendor benchmarks, but the early hands-on reports will settle that soon enough.

The license fine print

The license is derived from MIT and stays genuinely permissive: you can modify it, fine-tune it, and use it commercially. The one carve-out is hosted Model-as-a-Service offerings - you can't just wrap it in an API and sell access to the raw model.

For self-hosters, indie devs, and companies that want AI they actually own, that's about as good as flagship licensing gets.

The European angle

There's also something quietly significant about a frontier-class model trained in Europe with open weights, at a moment when the AI map is mostly drawn between the US and China. Mistral has become the default proof that a third path exists - and "Le Chonk" is its loudest argument yet.

One honest footnote before you clear disk space: a model this size is not a weekend laptop project. With llama.cpp support in place, most of us will realistically meet Le Chonk through quantized community builds rather than a full-precision home rig - and that's fine, that's how the open ecosystem has always worked.

If you've been looking for an excuse to build a serious local AI rig, a trillion open parameters is a pretty good one.

Sources: mistral.ai/news/mistral-large-4-oss/ huggingface.co/blog/mistral-large4 thenextweb.com/news/mistral-large-4-open-weig