What can you do with the Mistral Large 4 weights, and what does its licence forbid?

Mistral published the Large 4 weights on October 29, 2026: about one trillion parameters, 49 billion active per token, trained in Europe. You may use, change and sell them, and internal use is free at any size, but anyone running a model-as-a-service business needs a commercial licence from Mistral, whatever its revenue.

What can you do with the Mistral Large 4 weights, and what does…

Yesterday, Mistral AI released the weights of Mistral Large 4, nicknamed Le Chonk, three weeks after its public preview. Weights are the learned numbers of a model, and with them you can run it on your own hardware. This is not the same model as the preview. Mistral says those three weeks of further reinforcement learning brought large gains in coding and defensive security, and that it now uses fewer tokens.

For a European lab selling sovereignty, the release has a clear message: a large model trained from scratch in Europe on 3,800 NVIDIA Grace Blackwell chips in Mistral's own data centres, with a lineage you can trace.

What was released

Item	Mistral Large 4
Total parameters	About 1 trillion
Active per token	49 billion, or 52 billion with embeddings and output layers
Experts	128 routed, 4 active per token, plus 1 shared
Context window	1 million tokens
Inputs	Text and images
Languages	More than 160, including all official EU languages
Reasoning effort	Low, medium or high (the default)
Licence	Mistral Ecosystem Open-Weight License

A parameter is one learned number, and a token is a small piece of text. In a mixture-of-experts model, only part of it works on each token, so active parameters set the compute and total parameters set the memory. That is why you will see two figures in the sources: 49 billion and 52 billion describe the same model, counted differently.

What it takes to run

This one is big. vLLM's setup guide, written with serving detail, gives these figures:

Version	Download	Memory per GPU on 8 GPUs	Validated on
FP8 (8-bit)	About 1.06 TB	About 128 GiB	GB300, and two machines of 8 H200s; AMD MI350X
NVFP4 (mostly 4-bit)	About 610 GB	About 76 GiB	GB300 (one 4-GPU tray), B200
Quark MXFP4 (AMD)	About 578 GB	Fits 4 GPUs	MI350 Series

Here's the plain reading. A team needs a server with four to eight of the newest data-centre GPUs just to begin, and more for many users at once. A standard PC is well out of range. Community GGUF files for llama.cpp exist, along with Unsloth builds, and they lower the bar, but a model this size stays out of reach for ordinary home hardware.

Most people will use the hosted route: the mistral-large-4 model on Mistral's API, or the free demo gallery on Hugging Face in its first week.

How it compares

Mistral's claims are specific, and some are modest. On the Artificial Analysis Coding Agent Index, it reports 48 percent, effectively tied with GLM-5.3, while using 47 percent fewer tokens per task. On Terminal-Bench 4.0, it reports 41 percent, up 13 points from the preview, and says that leads the open models. In Surge's blind comparisons, it says graders preferred it to Kimi K3, GLM-5.3 and other open models, behind only Claude Opus 5.5.

It also concedes a weakness. On the Artificial Analysis Intelligence Index v5, it scores 26, the highest among Western open models, but its card says the leading Chinese open models score higher overall. The company argues its strengths are coding and security.

On security, it reports 51 percent on the Artificial Analysis Cyber Index and a 96.6 percent refusal rate on malicious cyber requests, which it says is the highest it measured among open models. Fine-tuning can weaken safeguards once weights are public, so treat this as a description of the model as shipped. The post's quotes come from companies working with Mistral, such as Booz Allen and Cisco, so read them in that light.

Try it: Mistral offers a bilingual angle few labs can match. Run five tasks in Québec French on the hosted model, as in our guide to trying Large 4, and compare them with the preview scores you kept, if any.

Why the weights matter in Canada

If you need data to stay under your own control, you now have another choice from outside the American and Chinese supply chains. A Canadian organization can run this on its own servers, and the model's French is a strong reason to test it for Québec. Nothing leaves your site when you self-host, though the chips, software and support still come from abroad. Weigh the cost honestly: at this size, self-hosting only makes sense for steady, high volumes or strict rules on data.

Do you need a commercial licence from Mistral?

Most organizations don't. The licence is a modified MIT that grants use, copying, change, sale, further training and derivative works. Here is the test:

Your situation	Commercial licence needed?
Internal use: nothing made available to third parties	No
End-user product with the model embedded in specific features	No
Relaying requests to a model hosted by someone else	No
Giving third parties API or endpoint access where they meaningfully control inputs, parameters or training data	Yes
You, or an affiliate, run such a business, even if this model isn't part of it	Yes, before commercial use

Compare this with its peers, and one detail stands out: there is no revenue threshold. Kimi K3's similar clause starts at $20 million US and GLM-5.3's at $10 billion US. A two-person startup selling a hosted API on Large 4 still needs a licence. The final row is also broad, since the clause covers a licensee "or any of its affiliates" that operates such a business, and that's a clause to show your lawyer.

The boundary between "embedded" and "meaningful control" won't always be clear. A support chatbot on your site is plainly a product. A platform where customers choose system prompts, parameters and their own training data looks much closer to a service. If you are near that line, use Mistral's contact form and get an answer in writing.

This is a change of terms for the line. Mistral Large 3 and Mistral Small 4 are tagged Apache 2.0 on Hugging Face and remain available, so a hosting business that needs that freedom can look there.