Mistral AI: history, people and positions

History

Three friends and a magnet link

Arthur Mensch, Guillaume Lample and Timothée Lacroix had known each other since their student days in the French engineering schools. By 2022 Mensch was at DeepMind's Paris office, where he co-wrote the Chinchilla paper on how to size language models, and Lample and Lacroix were at Meta's Paris AI lab, where both worked on the LLaMA models that Meta released to researchers in February 2023. Mistral says the company was born in April 2023. When it was four weeks old it raised a 105 million euro seed round led by Lightspeed Venture Partners at a 240 million euro valuation, with Xavier Niel, Eric Schmidt and the French state bank Bpifrance among the shareholders. Mensch told TechCrunch, "Open source is a core part of our DNA." Cédric O, a former French minister for digital affairs, became a non-executive co-founder and adviser.

On 27 September 2023 Mistral released its first model by posting a torrent magnet link on X, with no demo or launch video. Mistral 7B, a 7.3-billion-parameter model, beat Meta's Llama 2 13B on the company's benchmarks and came under the Apache 2.0 license, which allows any use. Its blog post said the chat version "does not have any moderation mechanism" and that Mistral looked forward to working with the community to make the model "finely respect guardrails." 404 Media, testing it with questions compiled by the safety researcher Paul Röttger, found that it would give instructions for suicide or for making crack; Röttger said safety had not even been mentioned in Mistral's public communications. In December 2023 it released Mixtral 8x7B, a sparse mixture-of-experts model, in the same way, and raised a 385 million euro Series A led by Andreessen Horowitz at a valuation of about 2 billion dollars.

The AI Act fight

Mistral was founded in the middle of the final negotiations on the EU AI Act. In June 2023 the European Parliament had voted for obligations on developers of foundation models; in November France, Germany and Italy pushed back, proposing self-regulation through codes of conduct instead. Mistral made its case in public. Mensch told TechCrunch that "the deployer should bear the risk, bear the responsibility," that hard product-safety rules on applications would push model makers to compete on safety, and that foundation models were like programming languages, which no one regulates directly. Negotiators reached a final deal in December 2023.

Critics said Mistral had helped weaken the law. Corporate Europe Observatory, a lobby watchdog, wrote in March 2024 that the developers of the most advanced systems had "scored a major victory" in that deal and that lobbying by Mistral and Germany's Aleph Alpha "tipped the scales," noting that Cédric O had called for more AI regulation while a minister and now had privileged access to the French government. It quoted Mensch telling Le Monde that in its final form the Act was "perfectly manageable for us." When Microsoft announced a 15 million euro investment in Mistral in February 2024, alongside putting Mistral Large on Azure, some members of the European Parliament said they were furious, arguing that the Act had been softened to help a European champion that was now tied to a US company; the Commission said it would look at the deal. A spokesperson for France's economy and digital ministry told Euronews that Microsoft's investment was small and that Mistral would remain independent.

In July 2025 Mensch was among the executives who signed the EU AI Champions letter asking the Commission to "stop the clock" on the Act for two years. The Commission refused. Its spokesperson said, "There is no stopping the clock. There is no grace period. There is no pause." In April 2026 Mistral published its own 22-point policy paper, which called for simplifying EU digital rules "without sacrificing underlying regulatory goals."

Open weights and paid models

Mistral has always sold some models and given others away, and the balance has shifted. In February 2024 its flagship, Mistral Large, was available only through its API and Microsoft Azure, which some developers read as a retreat from openness. In May 2024 it released the coding model Codestral with downloadable weights but under a new Non-Production License that barred commercial use without a paid license. Mensch told TIME in 2024 that nothing had changed: "We always had the intention to have leading models in the open source world, but also have some premium features that could only be accessed through monetized services."

In January 2025, with the release of Mistral Small 3, the company said it was "renewing our commitment" to Apache 2.0 for its general-purpose models and moving away from its restricted research license. In February 2025 it relaunched its Le Chat assistant, with apps for iOS and Android and paid tiers. In December 2025 it released the Mistral 3 family, including Mistral Large 3, a mixture-of-experts model with 675 billion total and 41 billion active parameters trained on 3,000 Nvidia H200 chips, all under Apache 2.0. Independent benchmarks placed Large 3 behind the closed models from Anthropic, Google and OpenAI on hard reasoning tests, and in July 2026 Mensch said a new, larger open-weight model was entering early access with partners.

From model maker to AI cloud

Mistral's funding grew with each round. In June 2024 General Catalyst led a 600 million euro Series B, part equity and part debt, valuing the company at about 6 billion dollars, and Mensch said the company remained "fully under the founders' control." In June 2025, at the VivaTech conference, Mistral announced Mistral Compute, a cloud business built on Nvidia hardware and offered to customers in Europe and elsewhere who wanted an alternative to US or Chinese providers; the company wrote that AI infrastructure "should not be restricted to two or three powerful companies." In September 2025 the Dutch chip-equipment maker ASML led a 1.7 billion euro Series C at an 11.7 billion euro valuation and agreed to work with Mistral on AI for its own products.

The infrastructure push accelerated in 2026. In February Mistral announced a 1.4 billion dollar investment in data centers in Sweden and bought Koyeb, a Paris serverless-computing start-up, its first acquisition, to build what Lacroix called "a true AI cloud." In July Microsoft and Mistral expanded their partnership with a multibillion-dollar agreement under which Mistral adds GPU capacity, including Nvidia Vera Rubin chips, for Microsoft's European customers, and Mistral models run inside Microsoft Foundry and Copilot Studio. On 8 September 2026 Samsung Electronics led a 3 billion euro Series D at a valuation of more than 21 billion euros, which Mistral called the largest equity round ever raised by a European technology company. Mensch told CNBC that the compute Mistral owns would roughly double over five years. TechCrunch reported that Mistral's decision to host other companies' open models, including Chinese ones, had led some observers to ask whether it was becoming a services provider; Mistral described its frontier research as "the foundation underpinning its infrastructure, products and sovereignty."

Timeline

Apr 2023

Mistral AI is founded in Paris

Arthur Mensch of DeepMind and Guillaume Lample and Timothée Lacroix of Meta start the company. Source
Jun 2023

105 million euro seed round

Lightspeed leads the round at a 240 million euro valuation, four weeks after the company's creation. Source
Sep 2023

Mistral 7B is released as a torrent

The first model comes out under Apache 2.0 through a magnet link posted on X, with no moderation built in. Source
Nov 2023

Mistral argues against AI Act rules for model makers

Mensch says deployers of AI applications, not foundation model developers, should bear the regulatory burden. Source
Dec 2023

Mixtral 8x7B and a 385 million euro Series A

An open mixture-of-experts model arrives the same day that Andreessen Horowitz leads a round valuing Mistral at about 2 billion dollars. Source
Feb 2024

Mistral Large and a Microsoft partnership

The API-only flagship launches with Azure as first distribution partner, and Microsoft invests 15 million euros. Source
Mar 2024

Lobby watchdog accuses Mistral of gutting the AI Act

Corporate Europe Observatory says Mistral and Aleph Alpha helped France and Germany dilute rules for general-purpose AI. Source
May 2024

Codestral ships under a non-production licence

The coding model's weights can be downloaded, but commercial use needs a paid license. Source
Jun 2024

600 million euro Series B

General Catalyst leads a round of equity and debt valuing Mistral at about 6 billion dollars. Source
Jan 2025

Mistral Small 3 renews the Apache 2.0 commitment

The company says its general-purpose models will be released under Apache 2.0 rather than its research license. Source
Jun 2025

Mistral Compute is announced

Mistral launches a European AI cloud built on Nvidia reference architectures. Source
Jul 2025

Mensch signs the "stop the clock" letter

European executives ask for a two-year pause of the AI Act; the Commission refuses. Source
Sep 2025

ASML leads a 1.7 billion euro Series C

The round values Mistral at 11.7 billion euros and starts an industrial partnership with the chip-equipment maker. Source
Dec 2025

Mistral 3 and Mistral Large 3

A 675-billion-parameter mixture-of-experts flagship and three small models are released under Apache 2.0. Source
Feb 2026

First acquisition, Koyeb

The Paris serverless start-up joins Mistral Compute days after a 1.4 billion dollar data-center commitment in Sweden. Source
Apr 2026

A policy playbook for European AI

Mistral publishes 22 recommendations on talent, regulation, procurement and compute. Source
Jul 2026

Microsoft partnership expands

A multibillion-dollar infrastructure agreement brings Mistral GPUs and models to Microsoft's European customers. Source
Sep 2026

Samsung leads a 3 billion euro Series D

The round values Mistral at more than 21 billion euros. Source

What it built

The main models and products to come out of Mistral AI, oldest first.

Sep 2023 · model

Mistral 7B

The company's first model, released under Apache 2.0 through a torrent link posted on X.

Source
Dec 2023 · model

Mixtral 8x7B

An open sparse mixture-of-experts model using 12.9 billion of its 46.7 billion parameters per token, which Mistral said matched or beat GPT-3.5 on most benchmarks.

Source
Feb 2024 · product

Le Chat

Mistral's chat assistant, opened as a beta running Mistral Large, Small and Next, and renamed Vibe in 2026.

Source
May 2024 · model

Codestral

A coding model with downloadable weights under a license that required payment for commercial use.

Source
Jun 2025 · product

Mistral Compute

A European AI cloud built on NVIDIA reference architectures.

Source
Dec 2025 · model

Mistral Large 3

A 675-billion-parameter mixture-of-experts flagship released under Apache 2.0 with three smaller Mistral 3 models.

Source
Mar 2026 · model

Mistral Small 4

A 119-billion-parameter Apache 2.0 model, 6 billion active per token, that merged Mistral's separate reasoning, vision and coding models into one.

Source

Where it stands

The lab's own stated positions, from its charters, policies and official statements.

Who should have access to powerful models?

Since January 2025 Mistral has committed to releasing its general-purpose models under the Apache 2.0 license, with weights free to download, modify and use, while keeping some specialist and premium models for paid use.

“We're renewing our commitment to using Apache 2.0 license for our general purpose models, as we progressively move away from MRL-licensed models.”

Mistral Small 3 (Mistral AI, January 2025), 2025 · Compare the people on this

Who should set the rules for AI?

Mistral's April 2026 policy paper asks the EU to simplify its digital rules, including the AI Act and GDPR, and to build a single compliance portal, while keeping the rules' aims.

“The complex EU digital regulatory framework should be revised to clarify inconsistencies, eliminate overlaps, and reduce compliance efforts without sacrificing underlying regulatory goals.”

European AI: a playbook to own it (Mistral AI, April 2026), 2026 · Compare the people on this

People

Here now

Arthur Mensch Brief profile
Co-founder and Chief Executive Officer, 2023-present

DeepMind co-author of the Chinchilla scaling paper who co-founded Mistral AI in 2023 and runs it as CEO.
Guillaume Lample Brief profile
Co-founder and Chief Science Officer, 2023-present

FAIR researcher behind unsupervised translation and the first LLaMA models, now Mistral AI's chief scientist.
Timothée Lacroix Brief profile
Co-founder and Chief Technology Officer, 2023-present

Meta engineer and LLaMA co-author who co-founded Mistral AI in 2023 and runs its engineering as CTO.

Sources

Last verified 2026-09-25.