Mistral AI

• 1. Introduction

  ›

  Mistral AI released its first language model on the 27th of September 2023, through a BitTorrent magnet link shared on the internet. The model, called Mistral 7B, was published under the Apache 2.0 license, making it free to download and use. Mistral had been a company for five months.

  The company was founded in April 2023 and headquartered in Paris. Three AI researchers had started it, having recently left positions at some of the most competitive AI research institutions in the world.

  The name came from the Occitan language of southern France. The mistral is a cold, powerful wind that sweeps through the French south.

  What drew three researchers to leave well-resourced institutions and start over in Paris? What choices made their models competitive against rivals with much larger teams and budgets? And how did a company that gave away much of its work for free build one of the largest technology valuations in Europe?

  747372128

• 2. Three Researchers, One School

  ›

  Arthur Mensch, Guillaume Lample, and Timothée Lacroix first met as students at École Polytechnique, France's most selective engineering school. Their careers then diverged before converging again at Mistral AI. Mensch built expertise in advanced AI systems at Google DeepMind. Lample and Lacroix both went to Meta Platforms, where they worked as specialists in large-scale language models.

  When the three came together to start the company, they brought knowledge of what it takes to build AI at the frontier of the field. Mensch understood advanced research systems; Lample and Lacroix knew what it demands to scale models in production at vast size.

  The company they launched would, within eight months of its founding, reach a valuation above €2 billion.

  67812

• 3. Open by Default

  ›

  Mixtral 8x7B arrived on the 9th of December 2023 via a BitTorrent link posted on Twitter. The model appeared on Hugging Face two days later, with a blog post accompanying the release.

  Mistral 7B had already come with a pointed benchmark claim: the model was said to outperform LLaMA 2 13B on every test, despite having fewer than half as many parameters. For Mixtral 8x7B, Mistral pushed that claim further. The company said the model beat both LLaMA 70B and GPT-3.5 on most benchmarks.

  In March 2024, Patronus AI published research comparing how often leading language models reproduced verbatim text from books protected under American copyright law. OpenAI's GPT-4 did so in 44 percent of cases. Mixtral came in at 22 percent. Meta AI's LLaMA-2 was at 10 percent, and Anthropic's Claude 2 at 8 percent.

  On the 10th of April 2024, Mixtral 8x22B was released via a BitTorrent link on Twitter, with a Hugging Face post following shortly after. The model contained 141 billion total parameters. Not all of Mistral's releases carried open weights: Mistral Large, launched on the 26th of February 2024, was a proprietary offering available on Microsoft Azure, with output in five languages and coding capabilities included.

  Mistral Small 4, released in March 2026, would take the same approach to a finer scale: 128 expert groups, with only 4 active per token.

  7270717576

• 4. Mixture of Experts

  ›

  The design that made Mixtral 8x7B stand apart was called a sparse mixture of experts. Rather than engaging all of a model's parameters for every token, this architecture divides the network into distinct groups called experts. Mixtral 8x7B had 8 such groups, giving it 46.7 billion usable parameters in total. Any single token only engaged 12.9 billion of those parameters at a time. That meant the model ran at the cost and speed of a 12.9-billion-parameter system while drawing on a much broader knowledge base.

  Mistral Large 2, announced on the 24th of July 2024, carried 123 billion parameters. The company released it with open weights under a Mistral Research License and claimed it was fluent in dozens of languages, including many programming languages. Its context length reached 128,000 tokens.

  On the 19th of November 2024, Mistral introduced Pixtral Large, which coupled a 1-billion-parameter visual encoder with Mistral Large 2.

  On the 2nd of December 2025, Mistral released Mistral Large 3, a sparse mixture-of-experts model with 675 billion total parameters and 41 billion active parameters. That same day, Ministral 3 arrived: three small, dense models with 3 billion, 7 billion, and 14 billion parameters, each with image understanding included.

  On the 17th of March 2025, Mistral released Mistral Small 3.1 with image understanding capabilities. The company described it as a new leader in the small models category.

  7175684583

• 5. From Seed to Billions

  ›

  In June 2023, two months after incorporating, Mistral closed its first funding round at €105 million ($117 million). Investors included Lightspeed Venture Partners, Eric Schmidt, Xavier Niel, and JCDecaux. The Financial Times estimated the company's valuation at €240 million ($267 million) at that point.

  On the 10th of December 2023, Mistral announced a second round raising €385 million ($428 million). Andreessen Horowitz, BNP Paribas, and Salesforce all participated, pushing the valuation past €2 billion.

  On the 26th of February 2024, Microsoft announced a $16 million investment in the company. By April 2024, reporting revealed Mistral was in talks to raise €500 million at a valuation of at least €5 billion. That round materialized: in June 2024, the company closed €600 million ($645 million) and reached a valuation of €5.8 billion ($6.2 billion). That figure placed Mistral fourth globally among AI companies by valuation, and first outside the San Francisco Bay Area.

  In April 2025, Mistral announced a €100 million partnership with the shipping company CMA CGM. In August 2025, the Financial Times reported that the company was in talks to raise $1 billion at a $10 billion valuation. Bloomberg then confirmed a €2 billion investment that valued Mistral at €12 billion. The Dutch semiconductor company ASML had contributed $1.5 billion of that total and taken an 11 percent ownership stake.

  In February 2026, Mistral acquired Koyeb, a Paris-based AI startup, and announced a multi-year partnership with Accenture. In March 2026, the company raised $830 million to build data centers near Paris and in Sweden.

  1213141719

• 6. Le Chat

  ›

  Le Chat, the French phrase for the cat, was the name Mistral gave its consumer AI assistant. On the 19th of November 2024, the company updated it with image creation capabilities, powered by Black Forest Labs' Flux Pro model.

  On the 6th of February 2025, Le Chat launched on iOS and Android. Mistral also introduced a Pro subscription tier at $14.99 per month, giving subscribers access to more advanced models, unlimited messaging, and web browsing.

  At the end of May 2026, the company renamed Le Chat as Vibe and introduced new features alongside the rebrand. Two months before the rename, in March 2026, Mistral had released Voxtral TTS, its first model designed for text-to-speech output.

  5831333236

• 7. Code, Math, and Reasoning

  ›

  Codestral, Mistral's first open-weight model built for code generation, launched on the 29th of May 2024. The company claimed it was fluent in more than 80 programming languages. Unlike many of Mistral's other open models, Codestral's license did not permit commercial use.

  A companion model, Codestral Mamba, used the Mamba 2 architecture, which supports longer input processing. It was released under the Apache 2.0 license, with only the instruction-tuned version made available.

  On the 16th of July 2024, Mistral released Mathstral 7B, a 7-billion-parameter model designed for STEM subjects. It was developed in collaboration with Project Numina and released under the Apache 2.0 license with a context length of 32,000 tokens.

  On the 10th of June 2025, Mistral released its first dedicated reasoning models: Magistral Small and Magistral Medium. Both were described as having chain-of-thought capabilities. Magistral Small was open-source; Magistral Medium was aimed at enterprise customers.

  On the 10th of December 2025, Devstral 2 and Devstral Small 2 arrived, both designed for software engineering tasks. Devstral Small 2, with 24 billion parameters, was claimed to outperform the Qwen 3 Coder Flash model, which carried 30 billion parameters, on coding benchmarks.

  Medium 3.5, released in April 2026, was described by Mistral as its first flagship merged model.

  6364596180