What Mistral AI builds
Mistral is a French company in Paris and the only large maker in this reference that ships its flagship with open weights: Large 3, 675 billion total parameters and 41 billion active, a 256k context window, under Apache 2.0. That means you can take it and run it on your own server.
- Flagship with open weights
- Four different licences
- A French company
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Model families
The model lines, from Large down to a 3B Ministral that fits on a laptop.
Versions
| Versions | Released | Status | Model families |
|---|---|---|---|
| Codestral 25.08 | current | Codestral | |
| Ministral 3 14B | current | Ministral 3 | |
| Ministral 3 3B | current | Ministral 3 | |
| Ministral 3 8B | current | Ministral 3 | |
| Mistral Large 3 | current | Mistral Large | |
| Mistral Medium 3.5 | current | Mistral Medium | |
| Mistral Small 4 | current | Mistral Small | |
| Voxtral TTS | current |
Why this maker is different for a reader in Iran
Every other maker in this reference publishes a supported-countries list, and Iran is on none of them. Mistral has an API too, and we have not tested it from Iran. But its three main models ship with open weights under Apache 2.0, and a model whose weights are published depends on no country list at all: you take the file and run it on your own hardware.
We do not write this as a way around sanctions, and it is not one. Apache 2.0 is a public grant the maker itself published. We are only saying that for someone in Iran looking for a model they can actually use, this difference matters more than any benchmark number.
Open means four things here, and that is where people get caught
Mistral is known as the open-model company, which makes people assume everything it publishes is free to use. It is not. In Mistral own list, Large 3, Small 4, all three Ministral 3 sizes, Shieldstral and Leanstral are Apache 2.0, so commercial use is fine. Medium 3.5 is under a modified MIT licence. Voxtral TTS is CC BY-NC 4.0, and the NC means no commercial use. Codestral, Embed and Moderation publish no weights at all and are Premier.
The practical version: if you build your product voice with Voxtral TTS and charge money for that product, you are in breach. It is one row in a documentation table and nothing else warns you.
The names changed
Until this year Mistral had two names everyone knew: Le Chat for the assistant and La Plateforme for the API. Neither appears in its documentation today. In their place are Vibe, a single agent for work and coding that runs on web, mobile, the terminal and the editor; Studio, the developer console and the API; and Admin, which holds the organisation, billing and SSO.
If you are reading something that talks about Le Chat, that text is out of date.
A small contradiction they printed themselves
Mistral has planted a flag on European data sovereignty and announced in-region inference and new European infrastructure on 11 August 2026. In its own model list there is also a Chinese model: GLM-5.2 from Z.ai, labelled open. We do not print this as a gotcha. We print it because it shows that sovereignty here is about where inference runs, not about where the model was built, and people routinely read those as the same thing.
What we cannot tell you
The price of a model. Mistral pricing page carries subscription tiers, a free plan with $10 a month in API credits, Pro at $14.99, Team at $24.99 per user with a $50 monthly minimum, but it does not print a per-million-token rate for any model. So the Mistral rows in our model list have an empty price column. Empty means we could not read it, not that it is free.
What it is good at
- A flagship with open weights under Apache 2.0, which is unique among the large makers in this reference
- The Ministral 3 line goes down to 3B parameters, which means it runs on ordinary hardware
- A European company with a registered office in Paris, which matters for a project with a data-region requirement
- It publishes its own model register with a release date for each, which few makers do
Where it falls short
- It does not publish a per-million-token rate for any model, so you cannot cost an API workload in advance.
- The licences are not uniform and one of them forbids commercial use. Read the row before you build on it.
- Persian is not among the languages named for Large 3. The list is explicitly not exhaustive, so we do not say it is unsupported; we say it is not named.
- We have not tested the API from Iran and found no supported-countries list we could cite either way.
Our take
If you want to host the model yourself, or you have a data-region requirement, Mistral is the first place we look. If you want the highest score in the coding table, it is not here and you should be looking at Anthropic.
Questions people actually ask
Does Mistral work from Iran
We have not tested the API and found no supported-countries list we could cite, so we say nothing about the hosted service. But the weights for Large 3, Small 4 and Ministral 3 are published under Apache 2.0, and a model with open weights runs on your own hardware with no account at all.
Are all Mistral models open
No. Large 3, Small 4 and Ministral 3 are Apache 2.0, Medium 3.5 is under a modified MIT licence, Voxtral TTS is CC BY-NC 4.0 which forbids commercial use, and Codestral publishes no weights at all.
What happened to Le Chat
The name no longer appears in Mistral documentation. The chat and coding agent is now called Vibe, and the developer console is Studio.