Mistral Large 4’s Licence in Plain English: Who Can Use It Commercially — and Who Needs to Pay

Table of Contents

• The One-Sentence Version
• What Mistral Actually Released on October 29
• Open-Weight Is Not Open-Source
• What the Licence Lets You Do for Free
• The One Condition: Model-as-a-Service
• What Counts as Model-as-a-Service — and What Doesn’t
• The Affiliate Clause: The Part Businesses Miss
• How Large 4’s Licence Compares With Mistral’s Earlier Models
• What It Costs to Actually Use It
• Why US Businesses Should Care
• Five Questions to Ask Before You Build on It
• Frequently Asked Questions
• Sources

---

When Mistral released the weights of Large 4 — nicknamed “Le Chonk” — on October 29, 2026, the headlines called it a trillion-parameter model anyone can download. That is true, and it is also the least interesting part of the story. The interesting part is the licence, because it decides what “anyone can download” actually means for a business.

The licence is called the Mistral Ecosystem Open-Weight License, and Mistral describes it as a modified MIT licence. For most people and most businesses, it is genuinely free: use it, modify it, fine-tune it, build products on it, sell those products. But there is one condition, and it is aimed squarely at one kind of business — companies that sell access to the model itself as a service. If that is your business, you need a separate commercial deal with Mistral first.

This guide reads the licence in plain English, explains the one condition and its grey areas, and compares it with what Mistral did before — so you can tell whether your plans are free, or whether you need to talk to Mistral first.

The One-Sentence Version

Mistral Large 4 is free to use, modify, fine-tune and commercialise — unless you or one of your affiliates runs a Model-as-a-Service business, in which case you must sign a separate commercial licence with Mistral before using the model or anything built from it for any commercial purpose.

What Mistral Actually Released on October 29

Large 4 spent three weeks as a public preview through Mistral’s API after its introduction on October 6. On October 29, Mistral published the full weights on Hugging Face, as announced on its blog, so anyone can download the model and run it on their own hardware instead of calling Mistral’s hosted service.

The model itself is a 1-trillion-parameter, natively multimodal mixture-of-experts model, with about 52 billion parameters active per token (49 billion plus embeddings and output layers), a 1-million-token context window, and fluency in more than 160 languages. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own data centres in Europe. It is built for coding, cybersecurity, agents and enterprise work, with three reasoning modes — low, medium and high — so the same model can handle a quick reply or a long agentic task.

Its benchmark results, as reported by Mistral, are strong for an open-weight model: 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than the leading open-weight rival, 19.7% on Cognition’s FrontierCode 2.0, 51% on the Artificial Analysis Cyber Index, and a 96.6% refusal rate on malicious cyber requests. Those are Mistral’s own numbers, not yet independently replicated.

For the full picture of what the model can do, see our Mistral Large 4 release explainer, which covers the cybersecurity angle in depth. This article is about the legal wrapper around it.

Open-Weight Is Not Open-Source

The first confusion to clear up: open-weight does not mean open-source. The weights — the trained numbers that make the model work — are available to download. The training data is not. The training code is not. Nobody outside Mistral can fully reproduce how the model was built.

That distinction matters for two reasons. First, it means you can inspect, run and adapt the model, but you cannot audit exactly what it was trained on. Second, it is why the licence is a bespoke document rather than a standard open-source licence. The question of what AI models are trained on is a live one — see our guide to the new robots.txt AI opt-out for how websites are responding.

What the Licence Lets You Do for Free

The licence grants permission, free of charge, to anyone who obtains a copy of the model to:

• Use it, copy it, modify it, merge it, publish it, distribute it, sublicense it and sell copies of it
• Run it, deploy it, further train it (fine-tune it) and create derivative works from it
• Let other people you give the model to do the same

There are two strings attached to all of that. The first is simple: you must include Mistral’s copyright notice and the permission notice in all copies or substantial portions of the model. That is the standard MIT-style attribution requirement, and it is easy to satisfy.

The second string is the one that matters, and it is the whole reason this licence exists.

The One Condition: Model-as-a-Service

Here is the licence’s second condition, in its own words (read the full licence text if you want every clause): “If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.”

Read that carefully, because it is stricter than it first looks. If the condition applies to you, it does not just cover the model-as-a-service part of your business. It says you need the commercial licence before using the model or its derivatives for any commercial purpose at all. The trigger is what your business does; the consequence covers everything commercial you do with the model.

Mistral’s reasoning is easy to understand. A trillion-parameter model is enormously expensive to train. Mistral is happy for businesses to build products on top of it — that grows its ecosystem. What it does not want is a cloud provider downloading the weights, wrapping them in an API, and selling access to Mistral’s own model back to the market without paying Mistral anything. That is the business the condition is designed to catch.

What Counts as Model-as-a-Service — and What Doesn’t

The licence defines “Model as a Service” as giving a third party access to the model or its derivatives for inference or post-training, by any means — for instance via an API or hosted endpoints — in a manner that allows that third party to exercise meaningful control over the inputs, parameters, or training data.

That “meaningful control” phrase does a lot of work. The licence then carves out two things that are not Model as a Service:

Situation	Model-as-a-Service?
Selling API access to Large 4 itself, where customers control the prompts, parameters and fine-tuning	Yes — you need a commercial licence
A cloud provider hosting Large 4 and charging per token	Yes — you need a commercial licence
Running Large 4 inside your own company, for your own employees	No — internal use is exempt
A product that uses Large 4 inside one specific feature (a coding assistant, a document summariser, a support chatbot)	No — model capability embedded within specific features is excluded
Simply passing a user’s request through to a model hosted by someone else	No — mere relaying is excluded
Using Large 4 through Mistral’s own API or products	No — use through Mistral’s official products is exempt

The practical line is this: if your customers are buying the model, you need a deal. If your customers are buying your product, and the model is one ingredient inside it, you almost certainly do not.

That distinction protects most real-world use. A law firm that fine-tunes Large 4 on its own documents and uses it internally is fine. A startup whose app uses Large 4 to draft contracts inside the app is fine. A company that launches “Large 4 as an API, 20% cheaper than Mistral’s” is not fine — and is exactly who the clause is for.

The Affiliate Clause: The Part Businesses Miss

There is one more detail in the condition, and it is the one most coverage has skipped over: the trigger applies if you or any of your affiliates operates a Model-as-a-Service business.

So consider a large company with many subsidiaries. One subsidiary runs a cloud platform that sells hosted AI models. Another subsidiary — a completely different team, in a different division — wants to use Large 4 inside its own product. On a strict reading of the licence, the first subsidiary’s business could trigger the commercial-licence requirement for the second subsidiary’s use, because the condition is written at the level of the licensee and its affiliates together.

There is a genuine grey area here: the licence does not define “affiliate.” In ordinary legal usage it means a company under common control, but the licence itself never says so, and the boundary matters a lot for big groups. If you work for a large organisation with a cloud or AI-hosting arm anywhere in the group, this is the question to put to your own lawyers — or to Mistral — before you build anything commercial on Large 4. Do not assume the condition stops at your team’s door.

How Large 4’s Licence Compares With Mistral’s Earlier Models

Mistral has not been consistent about licences, and that inconsistency is itself informative. Each model got the licence that suited Mistral’s strategy at the time:

Model	Licence	What it meant
Mistral 7B (Sep 2023)	Apache 2.0	Fully open, including commercial use, no MaaS condition
Large 2 (Jul 2024)	Research / non-commercial	Commercial self-deployment needed a separate licence
Large 3 (Dec 2025)	Apache 2.0	Fully open again
Large 4 (Oct 2026)	Mistral Ecosystem Open-Weight License	Free for almost everything, except Model-as-a-Service businesses

Large 4’s licence sits between the two extremes. It is far more permissive than Large 2’s research-only terms, and only slightly more restrictive than Apache 2.0 — with the restriction aimed at exactly one business model. That is a deliberate commercial choice: Mistral wants Large 4 everywhere except in a competitor’s API catalogue.

The comparison also matters if you are choosing between open models. Gemini 4 Carbon, Google’s current flagship, is closed — you can only use it through Google. Large 4 is the most capable Western open-weight model by Mistral’s own benchmark claims, and it is the one you can actually download. Our best AI models of 2026 comparison puts the three frontier options side by side.

What It Costs to Actually Use It

The licence is free, but running a trillion-parameter model is not free. You have two routes:

Use Mistral’s API. After the weights shipped, Mistral cut the API price roughly in half: about $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07. That is down from $1.36 and $4.18 during the preview. Using the API also sidesteps the licence question entirely, because use through Mistral’s official products is exempt from the MaaS condition.

Self-host it. The weights are on Hugging Face, with a quantized NVFP4 checkpoint for Blackwell GPUs, and deployment guides for vLLM and SGLang. Mistral’s own vLLM recipe uses eight GPUs in parallel, which tells you the scale of hardware involved — this is not a model you run on a laptop. The payoff is control: your data stays on your infrastructure, nobody can revoke your access, and you can fine-tune on your own data. For regulated US industries — finance, health care, defence — that control is often the whole point.

Why US Businesses Should Care

Mistral is French, and Large 4 was trained in Europe — but the licence questions land hardest on US readers, because the US is where the cloud providers, AI startups and enterprise buyers are.

For a US small business or developer, the news is good: you can download one of the world’s most capable open models, fine-tune it, and sell products built on it, without paying Mistral and without asking permission — as long as you are not reselling the model itself as a service. For a US cloud provider or API reseller, the news is a bill: if you want to offer Large 4 as a hosted service, you need to negotiate with Mistral first.

The open-weight frontier is now a race between Mistral, Chinese labs such as Moonshot, and others, while Google, OpenAI and Anthropic keep their flagships closed. Large 4 is the leading Western open-weight option, and its licence is the price of admission to that ecosystem. Understanding the licence before you build is cheaper than discovering the condition after you have shipped.

Five Questions to Ask Before You Build on It

1. Am I selling the model, or selling my product? If customers pay for your product and the model is embedded in a feature, you are on the free side of the line.
2. Does any affiliate of my company run a Model-as-a-Service business? If yes, get advice before using Large 4 commercially — the condition is written at group level.
3. Am I giving third parties meaningful control over inputs, parameters or training data? That is the licence’s own test for Model as a Service. A chatbot with fixed prompts is not that; a raw API is.
4. Do I need to self-host at all? If the API meets your needs, using Mistral’s official products avoids the licence question entirely.
5. Have I kept the attribution notice? Whatever you build, include Mistral’s copyright and permission notices in copies of the model. It is the one condition that applies to everyone.

Frequently Asked Questions

Sources

• Mistral AI — “Mistral Large 4 (Le Chonk) is now open weight,” October 29, 2026: https://mistral.ai/news/mistral-large-4-oss/
• Hugging Face — Mistral-Large-4-1T-A52B model card (licence summary, specifications, benchmarks): https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
• Mistral Ecosystem Open-Weight License (full text): https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE
• Artificial Analysis — Coding Agent Index and Intelligence Index v5 (as cited by Mistral)

General information only, not legal advice. Licence terms can change, and the affiliate question is genuinely unsettled — if your business model is anywhere near Model-as-a-Service, confirm your position with Mistral or your own lawyer before you build.