Mistral raised €3 billion on open weights. Read the licences before you ship.

Mistral AI has raised "€3 billion in a Series D funding round" at a "post-money valuation of more than €21 billion", which the company calls "the largest equity fundraising round ever completed by a European technology company". Samsung Electronics led it. The Scaleup Europe Fund, managed by EQT, and the existing investor PSG Equity co-led. Advent, funds and accounts managed by BlackRock, and the Grand Duchy of Luxembourg came in new. NVIDIA, ASML, a16z, Index Ventures, Lightspeed, General Catalyst, Salesforce Ventures and Bpifrance were already there and stayed.

The headline on Mistral's own announcement is about sovereignty and open weights. The company says it "operates across 20 countries" and "supports 125+ global enterprises", naming Airbus, ASML and HSBC among them.

If you are deciding whether to build on Mistral, none of that is the question you actually need answered. The question is narrower and much more practical: which of these models can I download, and what am I allowed to do with it once I have?

Mistral's documentation answers it. The answer is that "open weight" is doing a lot of work, because four different licences sit underneath it.

The licence map#

Every model below is one Mistral publishes weights for, with the licence its own model documentation lists against it and the API rate from its own pricing page.

Three things in that table are worth stopping on.

The first is good news, and it is the part of Mistral's pitch that survives contact with the documentation. Mistral Large 3 is Apache 2.0. Its flagship, the model its own docs describe as "A state-of-the-art, open-weight, general-purpose multimodal model", carries one of the most permissive licences in software. No revenue test, no field-of-use limit, no acceptable-use appendix that grows every quarter. Hugging Face's model API agrees: the repository is tagged license:apache-2.0. Most of the smaller line is Apache 2.0 too. That is a real commitment and it deserves to be said plainly, because plenty of companies use the words "open weight" for far less.

The second is Voxtral TTS, which is CC BY-NC 4.0. NC stands for non-commercial. If you were planning to put Mistral's text-to-speech into a product you charge for, that licence says no, and it says it regardless of how large or small your company is.

The third is the one that will catch people.

The $20 million clause#

Mistral Medium 3.5, the model its documentation calls "Our frontier-class multimodal model", is not Apache 2.0. It ships under something Mistral calls a Modified MIT License, and the modification is the whole story. From the licence file in the model's own repository:

You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month.

Read that carefully, because two details in it are easy to get wrong.

It is monthly, not annual. Twenty million dollars a month is roughly $240 million a year. This is not a clause aimed at startups, and if you have read elsewhere that Mistral caps its licence at $20m of revenue, that phrasing undersells the threshold by a factor of twelve.

It follows the model into whatever you build. The same clause says the restriction "applies to the Model and any derivatives, modifications, or combined works based on it, whether provided by Mistral AI or by a third party". Fine-tuning it does not get you out. Neither does obtaining it from someone else.

If you are over the line, the licence names your two options: contact Mistral at its sales address "to request a commercial license, which Mistral AI may grant you at its sole discretion", or use the model on Mistral's hosted service instead.

That is a defensible commercial design. It is close to what several other labs do, and a company that has just told its investors it is worth €21 billion has to make money somewhere. The problem is not the clause. The problem is that "open-weight" is the phrase attached to it in the marketing, and the two do not mean the same thing.

Hugging Face, notably, does not classify it as open either. Where Large 3 and Small 4 are tagged license:apache-2.0, Medium 3.5 is tagged license:other.

A small oddity in the drafting#

One detail is worth recording without over-reading it. The revenue clause refers to "This restriction in (b)". The licence's conditions are numbered 1 and 2. There is no (b) anywhere in the document.

It is almost certainly a leftover from an earlier draft, and the intent of the sentence is not in any real doubt. But it is in the operative clause of the licence governing Mistral's frontier model, and it is the sort of thing worth knowing about before your lawyers find it for you.

The model people actually pull is the restricted one#

Hugging Face publishes thirty-day download counts. Mistral Medium 3.5 recorded 123,804. Mistral Large 3, the Apache-licensed flagship, recorded 2,233.

The obvious explanation is size, not licensing. Medium 3.5 is a 128B model. Large 3 is 675B, which puts it out of reach of almost everyone who is not renting a rack. Our guide to how much VRAM you actually need to run a local LLM covers why that gap is so decisive in practice.

Still, the practical consequence stands: the Mistral model that most people can realistically download and run is the one with the revenue condition attached, and the one with the clean licence is the one most people cannot run.

The sovereignty pitch has a Chinese model on the price list#

One more thing sits on Mistral's API pricing page, listed under third-party models: GLM 5.2, from the Chinese lab Z.ai, at $1.40 per million input tokens and $4.40 output. Mistral's model documentation lists it too.

This is not hypocrisy and it should not be reported as such. Serving well-regarded third-party open models is ordinary platform behaviour, and GLM being available through a European provider is arguably a point in favour of the open-weight ecosystem Mistral says it wants. It is simply worth knowing that "sovereign AI", as the phrase is used in the announcement, describes where the company and its infrastructure sit rather than where every model on the menu was trained.

What to actually do#

If you are picking a Mistral model this week:

1. Under $20m a month in revenue? Everything in the table is available to you. Check Voxtral TTS separately, because the non-commercial licence binds you at any size.
2. Over it, or expecting to cross it? Mistral Large 3, Mistral Small 4 and the Ministral 3 family are Apache 2.0 and carry no threshold. If Medium 3.5 is the one you want, the decision is a commercial conversation with Mistral or a decision to stay on their API, and it is much cheaper to have that conversation before the model is embedded in a product.
3. Fine-tuning? The condition travels with the derivative. Assume the licence you started from is the licence you shipped.
4. Budgeting the API instead of self-hosting? Our back-of-envelope framework for what a model costs in production is the place to start, and remember that the self-hosting comparison has to include the hardware, not just the token rate. The real cost of GPU vendor lock-in is the part of that sum most teams leave out.

None of this is a reason to avoid Mistral. Apache 2.0 on a frontier-class flagship is genuinely rare and genuinely valuable, and it is more than several better-funded labs offer. It is a reason to read the licence file rather than the launch post, which is the same advice we gave when Mistral turned on training from user chats by default, and the same reason we listed the licence alongside the file size when DeepSeek V4 shipped its open weights.

"Open weight" is not one thing. On Mistral's own documentation it is currently four.