Mistral AI, the French lab behind open and premier models

A Paris startup born in 2023

Mistral AI was set up in April 2023 by three French researchers: Arthur Mensch, Guillaume Lample and Timothée Lacroix. Its stated goal is to make frontier AI “open to all” and to build a leading European player that combines innovation with openness, transparency and cost efficiency.

In September 2023 the company released its first model, Mistral 7B, under the Apache 2.0 license, which immediately put it on developers’ radar. It has since broadened its catalog and now works with businesses and governments in finance, manufacturing, energy, defense and the public sector.

A catalog split between open and premier

Mistral’s documentation draws a clear line between two categories:

• Open-weight models, mostly under Apache 2.0 (a few under a modified MIT license): general-purpose Mistral Small and Large models from some generations, compact Ministral models for modest hardware, and Voxtral audio models.
• Premier models, available only through the API: the latest versions of Mistral Large and Medium, Codestral for code, and OCR models for reading documents.

The family names are telling: Ministral for small language models, Codestral and Devstral for programming, Magistral for reasoning, Voxtral for speech.

How to use Mistral models

• Le Chat, the consumer assistant: see the Le Chat page.
• Mistral’s API and console, for developers, with detailed documentation.
• Downloading the weights of the open models, to host them on your own servers.

What sets Mistral AI apart

Its main selling point is sovereignty: a European company, subject to the GDPR, whose models you can in some cases run on your own infrastructure without relying on a US or Chinese provider. The trade-off is a tighter catalog than Google’s or OpenAI’s. To compare its open models with those from Meta, Alibaba or DeepSeek, see the open source models page.