Aiodic

Back to the table

61 Model family Mi Mistral Model Families

Mistral

Mistral AI's model family.

What it is

Mistral is a family of AI language models made by Mistral AI, a company based in Paris. It is the best-known AI model maker in Europe, and it is unusual in one way: it gives away some of its models for anyone to download and run, while selling others as paid services. That mix is what the popup on this table means by "open-weight and commercial".

The company was founded on 28 April 2023 by three French researchers who had met as students at the École polytechnique, one of France's top engineering schools. Arthur Mensch, who became chief executive, had worked at Google DeepMind. Guillaume Lample and Timothée Lacroix had both worked at Meta's AI research lab, and both were among the authors of the February 2023 paper that introduced Meta's first Llama model. In June 2023, less than two months after it was founded and with no product yet, Mistral raised €105 million, widely reported at the time as Europe's largest ever seed round, the first money a start-up raises.

What "open weights" means. A model's weights are its learned numbers, the billions of settings that hold everything it picked up in training (see Parameters). If a company publishes the weights, you can download the model and run it on your own computer or server, with no company in between. That is different from "open source" in the full sense, where you would also get the training data and the recipe. Mistral, like most labs, publishes weights but not its training data.

Three kinds of licence. A licence is the legal permission that comes with the download, and Mistral has used several. This is the part people most often get wrong, so it is worth slowing down:

Then there are models whose weights were never published for download, such as the first Mistral Large in February 2024. Mistral offered it through its own service and cloud partners, with self-hosting only by special arrangement for business customers. So "Mistral is open" is only partly true. Always read the licence of the exact model you want.

Where you meet it. Most people use Mistral through its chat app, which launched in February 2024 as Le Chat. In late May 2026 Mistral renamed it Vibe and combined it with its coding assistant; Mistral's help centre says old Le Chat accounts and conversations carried over and the chat.mistral.ai address still leads there. Developers use Mistral through an API, a way for one program to use another, paid per token, and through cloud services from Microsoft, Amazon and Google. Others simply download the open models.

At the time of writing, Mistral's model list includes Mistral Large 3, Mistral Small 4 and the small Ministral 3 models under Apache 2.0, and Mistral Medium 3.5 under its modified licence, alongside speech, document-reading and moderation models. The line-up changes every few months, so check the current list yourself.

Why it matters

It gives you a choice that isn't American or Chinese. Most of the leading model families come from US companies (GPT, Claude, Gemini, Llama) or Chinese ones (DeepSeek, Qwen). Mistral presents itself as a European alternative. In the company's framing, and increasingly in its investors' and French politicians' framing, this is about "sovereignty": the idea that European governments and businesses should be able to use AI without depending on a foreign company or keeping their data under foreign law. Whether a company that also partners with Microsoft and Nvidia delivers that independence is a fair question, and one this page returns to. But the demand is real. TechCrunch, reporting on Mistral's September 2026 funding round, cited Financial Times reporting that being non-American had helped Mistral's sales.

Its open models let you keep your data at home. If you run an open-weight model on your own computer, your words never leave the machine. For a therapist's notes, a lawyer's drafts, or an unpublished manuscript, that can matter more than having the most capable model available.

It is strong in European languages. When Mistral launched Mistral Large in February 2024, it said the model was natively fluent in English, French, Spanish, German and Italian. That is the company's claim, and you should test it in your own languages, but it is one reason many European users pick it for translation and multilingual writing.

It shows how open releases shape the whole industry. Mistral's early downloadable models set a bar that others had to meet, and its December 2023 Mixtral release put the mixture of experts design into ordinary developers' hands (that story is told on the Mixture of Experts page). Open releases also bring real risks, covered below.

Real-world examples

A model released as a torrent link (September 2023)

On 27 September 2023, Mistral released its first model, Mistral 7B, in an unusual way. Instead of a launch event or a press release, the company posted a single link on X: a magnet link, the kind used to download files by BitTorrent, a file-sharing method that spreads a download across many computers. The file contained the model's weights. VentureBeat described it as bucking the release trend of the time. A blog post and a download on Hugging Face, the main website for sharing AI models, followed.

Mistral 7B has about 7.3 billion parameters, small enough to run on a good home computer. Mistral said it outperformed Meta's Llama 2 13B, a model almost twice its size, on all the benchmarks it tested. It came under Apache 2.0, which meant anyone could use it in a business with no fee. It quickly became one of the most widely used open models for experiments and small products.

The release was also criticised. Within days, 404 Media reported that the model would give detailed answers to requests for harmful instructions, and quoted AI safety researcher Paul Röttger saying his biggest issue was that safety was not evaluated or even mentioned in Mistral's public communications. A torrent also cannot be taken back once it is shared. Mistral then added a note to its release page saying the model has no moderation mechanism, and said it looked forward to working with the community on ways to make it respect guardrails. The episode shows both sides of open weights in one story: freedom for anyone to build, and no way to recall a model after the fact.

The AI Act lobbying and the Microsoft deal (November 2023 to May 2024)

In autumn 2023 the European Union was finishing the AI Act, its law on artificial intelligence. The European Parliament wanted rules for "foundation models", the large general-purpose models that other products are built on. Mistral argued against this. In a November 2023 interview with TechCrunch, Arthur Mensch said the rules should target the companies that deploy AI in products, not the makers of the underlying models. Cédric O, a former French minister for digital affairs and by then a Mistral co-founder and adviser, helped organise an open letter from business leaders warning against heavy regulation and, according to TIME, warned that the law could "kill" Mistral. In November 2023, France, Germany and Italy proposed that foundation model makers should instead regulate themselves through codes of conduct. Critics, including members of the European Parliament and AI researcher Yoshua Bengio, strongly opposed that idea, and the Ada Lovelace Institute, a UK research body, said it would be irresponsible to drop regulation to protect "one or two national champions".

The EU reached a political deal on 8 and 9 December 2023. It kept obligations for general-purpose models, with extra duties for the most powerful ones, and lighter treatment for models released under free and open-source licences, unless they pose a systemic risk.

Then, on 26 February 2024, Microsoft announced a partnership with Mistral. It made the new Mistral Large available on its Azure cloud and included a small investment, reported at about €15 million. Some EU lawmakers reacted angrily. A senior adviser to MEP Axel Voss told Euronews they were "extremely furious", because the French government had argued for months that the AI Act must not hold back European leaders. The European Commission said it was looking at the deal, and Green MEPs asked it to examine possible conflicts of interest. The French government said Mistral had always been open about its business model. In May 2024, the UK's competition regulator decided that the partnership did not give Microsoft enough influence to count as a merger and closed its review.

People still disagree about this episode. A March 2024 report by Corporate Europe Observatory, a lobbying watchdog, argued that European start-ups had helped Big Tech weaken the law. Mistral's position was that rules aimed at model makers would hurt small European companies more than US giants. What is not in dispute is that the "European champion" argument carried weight in Brussels, and that the Microsoft deal made people question it.

Le Chat, a president's endorsement, and a rename (2024 to 2026)

Le Chat launched as a free test version on 26 February 2024, the same day as Mistral Large. At launch Mistral said it could not search the internet and might give outdated or inaccurate answers. Phone apps followed on 6 February 2025. Just before the AI Action Summit in Paris that month, French President Emmanuel Macron told viewers in a TV interview to download Le Chat rather than ChatGPT, as TechCrunch reported. Mistral told the French newspaper Le Parisien the app passed one million downloads within about two weeks.

In late May 2026 Mistral renamed Le Chat as Vibe and merged it with its coding tool, turning the chat app into what the company calls an agent that can carry out multi-step tasks across your apps. That makes it closer to the coding agents and agents described elsewhere on this table. At the time of writing there is a free plan and paid plans; heise reported that Mistral described some paid limits only as multiples of the free allowance, so check what each plan actually includes.

The money behind "European sovereignty" (2023 to 2026)

Mistral's funding tells the story of how investors see it. After the €105 million seed round in June 2023, it raised €385 million in December 2023 and €600 million in June 2024. On 9 September 2025, Mistral announced a €1.7 billion round led by ASML, the Dutch maker of the machines used to produce advanced computer chips, valuing Mistral at €11.7 billion. ASML put in about €1.3 billion for a stake of around 11 percent, and the two companies announced a partnership to use Mistral's AI in ASML's products and operations. On 8 September 2026, Mistral announced a €3 billion round led by Samsung Electronics, at a valuation reported at €21 billion, with the EU-backed Scaleup Europe Fund among the co-leaders.

Notice the pattern. The money increasingly comes from large industrial companies and public-backed funds that want AI they can control, which fits Mistral's sovereignty pitch. But investors in earlier rounds included US firms such as Andreessen Horowitz and Nvidia, and Samsung is Korean. "European" describes where Mistral is based and who it sells to, not that every euro or chip is European.

Try it yourself

Two short exercises. The first needs only a web browser. The second needs a reasonably modern computer and about 5 GB of free disk space. Menus and names change often, so treat the labels as a guide.

Exercise 1: test Mistral's multilingual claim

  1. Go to chat.mistral.ai and sign in or create a free account. At the time of writing the app is called Vibe (formerly Le Chat) and has a free plan with usage limits; check the current plan details.
  2. Pick a language you know besides English, or one a friend or customer speaks. Paste this prompt, filling in the brackets: Write a 120-word welcome message for customers of my [type of business]. Write it first in English, then in [second language]. Don't translate word for word: make the second version sound natural to a native speaker, and tell me in one line which phrases you changed and why.
  3. Check the second version. If you are not fluent, ask a native speaker, or paste it into a different AI tool and ask for honest feedback. Look for stiff phrasing, wrong levels of formality (for example "tu" versus "vous" in French), and cultural mismatches.
  4. Ask a follow-up to see how it revises: Make the [second language] version more formal, suitable for a bank or a law firm, and keep it under 100 words.
  5. Compare with another assistant. Paste the first prompt into ChatGPT, Claude or Gemini and judge which result a native speaker would prefer. Don't trust anyone's rankings, including the company's own.

Exercise 2: run Mistral 7B on your own computer

  1. Install Ollama from ollama.com. It is a free program that downloads open models and runs them on your computer. Install it like any other app.
  2. Open a terminal (Terminal on a Mac, PowerShell on Windows) and type: ollama run mistral The first time, this downloads Mistral 7B, about 4.4 GB at the time of writing. When a prompt appears, the model is running entirely on your machine.
  3. Try a private task, such as tidying rough notes. Type: Turn these notes into a clear, polite email to a supplier: [paste your rough notes]
  4. Test it offline. Turn off your Wi-Fi and ask another question. It still answers, which proves nothing is being sent anywhere.
  5. Notice the limits. Ask it about something recent, or a tricky fact. A small 2023 model knows less and makes more mistakes than the big online assistants, so check anything important. Type /bye to quit. For more on why model size matters, see Parameters.

Common mistakes and misunderstandings

Assuming every Mistral model is free to use in business. Only some are. Apache 2.0 models can be used commercially. Research and non-production licences forbid it without a paid licence, and the modified MIT licence excludes large companies. Check the licence for the exact model and version, not the family name.

Confusing open weights with open source. You can download and run the models, but Mistral does not publish its full training data or methods. You can't see exactly what the model learned from, and you can't rebuild it yourself.

Assuming open means safer or less safe. Neither is automatic. Open weights let independent researchers inspect and test a model, which is good. But anyone can remove the safety training from a downloaded model, and Mistral's first release shipped with no moderation at all. Safety depends on how a model is built and used.

Taking "European" to mean your data stays in Europe. Where your data is stored depends on the product, the plan and the settings, not on where the company is based. If this matters to you, read the current privacy terms and data settings for the plan you use, or run an open model on your own hardware.

Treating the sovereignty story as settled fact. It is Mistral's framing, and one shared by the French government and some investors. Critics point to the Microsoft partnership, the reliance on US-made chips, and the AI Act lobbying. Weigh both views.

Looking for "Le Chat" and not finding it. Since late May 2026 the app is called Vibe. Old guides and videos still use the old name.

Assuming the big online model is always the best choice. For simple, private jobs, a small open Mistral model on your laptop may be enough and costs nothing to run. For hard reasoning or current information, a larger hosted model will usually do better. Match the model to the task.

Forgetting that features and prices change. Model names, plans and licences have changed many times since 2023. Everything on this page marked "at the time of writing" should be checked against Mistral's own pages before you rely on it.

Related elements