Mistral's new open model has a trillion parameters — and it was trained in Europe

Mistral has pushed Mistral Large 4 into preview — a 1-trillion-parameter model the French company calls the strongest open model built outside China. It was trained from scratch on 3,800 Grace Blackwell accelerators in Mistral's own European data centers, and the weights are promised for release by the end of October. For now the model lives only behind an API: you can check the claim that it caught up with closed models, but you cannot take it home yet.

What was actually announced

Mistral Large 4 has 1 trillion parameters, of which 52 billion are active on any given request. That makes it a sparse model: fast to run, heavy to store. It is natively multimodal, so it handles images as well as text. Training took two months on 3,800 Grace Blackwell accelerators in Mistral's European data centers, and the company stresses that it trained the model from scratch rather than distilling someone else's outputs.

This is a preview. Access is through the Mistral Studio API, and the weights are promised "by the end of the month." Until then, Mistral is red-teaming the model in real settings with cybersecurity teams, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities.

Why anyone cares

For the past couple of years the strongest open weights came out of Chinese labs while Western companies played catch-up. Le Chonk is positioned as the answer: the most capable open model built outside China and "very, very close" to proprietary ones. The bet on openness here is practical rather than ideological — an open model runs on your own hardware and costs exactly what that hardware costs. In cybersecurity that matters: if a provider refuses to answer a question about a vulnerability, the work simply cannot be done, and losing access to a tool mid-incident is expensive.

What independent tests say

Official benchmarks are always a shop window, so third-party numbers are more interesting. On the Artificial Analysis aggregator the model scored 38. That is a big jump from last year's Mistral Large 3, which barely managed 9, but it still trails DeepSeek 4.1 Flash with its 552 billion parameters. In plain terms, Mistral is back in the game and roughly six months behind the frontier rather than years. For an open model you can host yourself, that is not bad at all.

Where the catch is

First: there are no weights yet. All you get is a preview API, and "by the end of the month" can slip. Second: a trillion parameters will not run on a home GPU. Even with 52 billion active parameters you need a serious cluster, so this openness is aimed at companies rather than hobbyists. Third: the model offers only two reasoning levels, "none" and "high," with nothing in between. And fourth, the honest one: a score of 38 is not the frontier. If you need maximum quality right now, an open model will not give it to you.

What it means in practice

The real story is not this one model but the fact that the gap between open and closed models keeps narrowing. You can see it not only in benchmarks but in how people actually pick models for work. According to our service's data over 30 days (window 08.09–08.10.2026, 757 messages from 74 users), the most requested chat model was by no means the heaviest: the lightweight gpt-5-4-mini accounted for 127 messages, while the larger gpt-6-astra took 91 and claude-opus-5 took 67. Images follow the same logic: in the generation_costs table for the same 30 days, the cheapest model, z-image, cost $0.004 per image (128 generations), while the popular nano-banana-2 cost $0.052 (2,206 generations) — thirteen times more. People and companies vote for good-enough quality at a sane price, not for the top score in a table.

FAQ

What is Mistral Large 4, and why is it called Le Chonk?

It is the new flagship model from France's Mistral, with 1 trillion parameters. Le Chonk is the unofficial nickname the company itself uses in the announcement; the official name is Mistral Large 4, or ML4 for short.

Can I download it yet?

No. Only a preview through the Mistral Studio API is available today. The open weights are promised by the end of October 2026.

How is it different from closed models?

An open model can be downloaded, fine-tuned and run on your own hardware without depending on someone else's API. Closed models cannot be used that way — access is only through the provider's service.

Where can I try modern models?

You can talk to a range of language models in NeuralSpace Chat, and build your own AI project in the Code section.