Names, versions and owners change fast. Checked October 2026.
In plain words
Mistral AI is a company, founded in Paris in April 2023 by three French researchers who had worked at Google DeepMind and Meta. It makes large language models in several sizes, for chat, coding, reasoning, speech and documents.
Mistral is known for releasing many of its models with open weights: anyone can download them and run them on their own computers or servers. Some specialised models stay closed and are offered only through its app and API.
Its chat app launched in February 2024 as Le Chat and was relaunched under a new name in May 2026. The box above has the app's current name and Mistral's newest models.
Everyday example
A bakery that also shares recipes
Picture a bakery that sells bread over the counter, but also prints most of its recipes for anyone to bake at home. A few special cakes stay secret and are only sold in the shop. Mistral works like that: many of its models can be downloaded and "baked" on your own machines, while a few are available only through its own services.
Where the example breaks: Recipes cost nothing to copy and bake. Running a large model at home needs serious computing power, so in practice many people still use Mistral's models through its own app or a cloud provider.
Latestas of October 2026
- Newest flagship
- Mistral Large 4, announced in preview 6 October 2026, about 1 trillion parameters with 52 billion active; its open weights, nicknamed Le Chonk, followed on 29 October 2026
- Chat app
- Vibe, the new name for Le Chat since May 2026
- Other recent models
- Voxtral 3 for speech (20 October 2026)
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
Mistral's models use the same broad recipe as other language models, with a few design choices it is known for.
Small but strong models
Its first model, Mistral 7B (September 2023), had 7 billion parameters, small by industry standards, and was released under the permissive Apache 2.0 licence.
Mixture of experts
From Mixtral (December 2023) onward, many large Mistral models use a mixture of experts: only part of the model runs for each token. Mistral Large 4, from October 2026, has about 1 trillion parameters, with about 52 billion active at a time.
A family for different jobs
Beyond general Small, Medium and Large models, there are Devstral for code, Magistral for step-by-step reasoning, and Voxtral for speech.
Run where you want
Because many weights are public, companies can run Mistral models inside their own data centres or in Europe, which matters to organisations with strict data rules.
- Company
- Mistral AI, Paris, founded April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix
- First model
- Mistral 7B, 27 September 2023, Apache 2.0 licence
- Known for
- Open-weight models, many under Apache 2.0, that companies can run on their own servers or in Europe
- Chat app
- Le Chat (February 2024), renamed Vibe in May 2026
Questions people ask
Is Mistral open source?
Many Mistral models are released with open weights under licences such as Apache 2.0, so you can download, use and modify them. Some, such as the Magistral Medium reasoning model launched in 2025, were offered only through Mistral's app and API. Training data is not published, so "open weights" is the more exact term.
What is Le Chat?
Le Chat was Mistral's chatbot app, launched in February 2024. In May 2026 Mistral relaunched it under the name Vibe, with more features for working on tasks and code. It runs on Mistral's own models.
Is Mistral the same as ChatGPT?
Who owns Mistral AI?
It was founded by Arthur Mensch, Guillaume Lample and Timothée Lacroix and is backed by outside investors. The Dutch chip-equipment maker ASML took a stake in September 2025.