In just three years, the French company Mistral AI has risen to become one of the names being discussed both in its native Europe and across the rest of the world. It builds large language models, offers some of them as open-source software, and promises to put advanced artificial intelligence into everyone’s hands. Yet many people misunderstand the company. Anyone expecting a “European equivalent of OpenAI” will probably be disappointed. Mistral is playing a different game than it might seem at first glance.
Attention surrounding the company has also grown because of developments across the Atlantic. After the U.S. administration forced Anthropic to withdraw its most powerful models and calls for technological independence intensified in Europe, Mistral found itself in the spotlight. But the French darling is also shrouded in numerous misconceptions, and the fact that it builds language models only makes the overall picture more confusing.
Three researchers and a wind from southern France
Mistral was founded by three men who met while studying at the École Polytechnique in Paris. CEO Arthur Mensch previously worked at Google DeepMind, while CTO Timothée Lacroix and chief scientist Guillaume Lample came from Meta. They founded the company in April 2023 and named it after the mistral, a powerful northwesterly wind that sweeps from southern France toward the Mediterranean Sea.
Mensch is not just a developer. He has become a public advocate for a particular vision of artificial intelligence and has managed to gain access to places that technology executives usually struggle to reach. Speaking before French lawmakers, he warned that Europe had two years left before definitively losing the AI race. Yet he himself admits that he still needs to properly explain his own company to people.
In an extensive post on LinkedIn, he described what the Paris-based company actually does. It deploys its models and agents directly on the infrastructure of major customers and helps them build custom models using the Forge platform, which can train on their own data. Mensch wants everyone to have access to the best artificial intelligence systems, beyond the oversight of governments and companies that need to control everything.
Europe’s equivalent of Palantir, not OpenAI
This is where the main misunderstanding lies. While observers compare Mistral with American giants and search for a French ChatGPT, the company is heading in a different direction. Its Vibe chat and agent tool, previously known as Le Chat, has only a fraction of ChatGPT’s public recognition. And Mistral is not even the favorite among startup founders at Paris’s Station F campus, where more people praise competing models.
Instead, the company is betting on a model that more closely resembles America’s Palantir. It sends its engineers directly to customers—governments and major corporations—and helps them implement artificial intelligence and adapt it to their specific needs. Mistral is thus positioning itself as a partner for demanding sectors, ranging from finance and manufacturing to defense, energy, and public administration.
The results reflect this approach. In February, the company revealed that its annual recurring revenue had exceeded $400 million, compared with just $20 million a year earlier. According to the company, it is targeting $1 billion this year.
Custom models, including for phones
Mistral has built a broad range of models. These range from conventional language models to multimodal and analytical systems, as well as models for audio and document processing. Not all of them rely on enormous scale. The company released the tellingly named Mistral Small 4 and the “Les Ministraux” family, models optimized for smartphones and other edge devices. It also released the Leanstral coding agent as open-source software.
Mistral had already attracted attention with so-called sparse mixture-of-experts models, which were able to match the performance of significantly larger rivals. It offers many of its models with open weights under the Apache 2.0 license, although its most powerful models are subject to certain restrictions on commercial deployment.
Another model is on the horizon. Mensch promised that it would arrive in the summer, have open weights, and become available for early access as soon as July. In areas that are less computationally demanding—voice, images, and document processing—the company remains at the forefront, according to Mensch.
Data centers and chips
At the beginning of the year, Mistral acquired infrastructure startup Koyeb, its first-ever acquisition, bringing it closer to its plan of building genuine infrastructure for artificial intelligence. It also added Austria’s Emmi, which specializes in physical AI and is expected to help support industrial customers.
At the same time, the company announced an investment of roughly €4 billion in data centers in France and Sweden. The undertone of technological self-sufficiency is unmistakable. Mensch presents artificial intelligence as a standard technology that every organization needs to have available securely and affordably.
The question of proprietary chips remains open. For now, Mistral relies on Nvidia, which Mensch describes as a great partner. However, he does not rule out developing the company’s own silicon. According to him, it should happen one day, but for now the company is merely testing a few things here and there.
From the French military to IBM
Mistral’s list of partners reads like a showcase of European industry and the public sector. In 2024, it signed an agreement with Microsoft that included a €15 million investment and distribution of its models through the Azure platform. This was followed by participation in building a Paris AI complex alongside MGX, Nvidia, and French state-owned bank Bpifrance.
At the VivaTech trade show, the company then announced Mistral Compute, a European platform built on Nvidia processors, which President Emmanuel Macron described as historic. He shared the stage with Mensch and Nvidia CEO Jensen Huang. In July 2025, the AI for Citizens initiative followed, intended to help governments and public institutions transform public services.
Mistral has also entered into partnerships with semiconductor equipment manufacturer ASML, consulting firm Accenture, news agency AFP, the French military and employment agency, Luxembourg, shipping giant CMA CGM, German defense startup Helsing, as well as Orange and automaker Stellantis. IBM holds a special place, having made the flagship Mistral Large 2 model available on its watsonx platform and enabled companies to run it on their own servers, with support for more than 80 programming languages and a context window of up to 128,000 tokens.
Mistral and funding
The company’s financing has one unusual feature. Mistral raised most of its money through debt, but altogether it has secured around $4 billion. Its start was extraordinary. Just one month after it was founded, in June 2023, the company closed a record $113 million funding round led by Lightspeed. At the time, it was the largest seed investment in European history, valuing the startup at $260 million.
Six months later, another round brought in €385 million, equivalent to $415 million at the time, led by Andreessen Horowitz at a company valuation of around $2 billion. Microsoft added a $16.3 million convertible investment in early 2024. In June of that year, Mistral raised €600 million through a combination of equity and debt in a General Catalyst-led round at a valuation of $6 billion, with Cisco, IBM, Nvidia, and Samsung also participating.
The company closed its most recent major funding round in September 2025. Led by ASML, it raised a total of €1.7 billion, or roughly $2 billion, and boosted Mistral’s valuation to approximately $13.8 billion. According to speculation, it is now preparing another round of around $3.5 billion, which would push the company’s valuation toward $23 billion. Even that is still less than what the leading American labs have at their disposal.
When asked what a potential exit from private ownership might look like, Mensch gave a clear answer in Davos in January 2025. The company is not for sale, and the plan is to go public.
Sources: techcrunch.com, mindstudio.ai, and mistral.ai
