Join the reliable event by business leaders for almost two decades. VB Transform brings together people who build a strategy of real business. Get more information
Mistral AI, the French artificial intelligence startup, on Wednesday announced a radical expansion in AI infrastructure that positions the company as Europe’s response to giants of computing in the US cloud, while simultaneously scares new rasons.
The company based in Paris revealed Mistral Compute, a comprehensive infrastructure platform built in association with NVIDIA, designed to provide European companies and governments with an alternative to trust the cloud based in the United States and Google. The movement repeats a significant strategic change for Mistral from the models of purely developing until the control of the entire technology stack.
“This movement towards the infrastructure of AI marks a transformative step for the Mistral, since it allows us to address a vertical criticism of the AI value chain,” said Arthur Mensch, CEO and co -founder of the Mistral. “With this change comes the responsibility of ensuring that our solutions not only promote innovation and adoption of AI, but also defend Europe’s technological autonomy and contribute to sustainability leadership.”
How Mistral built reasoning models that think of any language
Together with the infrastructure announcement, Mistral announced its master series or AI model systems for Reas, logical thinking step by step similar to the OPENAI O1 model and the Deepseek R1 of China. But Guillaume Lample, the main scientist of Mistral, says that the company’s approach differs from competitors crucially.
“We did everything from scratch, basically because we wanted to learn the experience we have, like, flexibility in what we do,” Lampra told me in an exclusive interview. “We are really achieved to be really very efficient in the stronger online reinforcement learning portfolio.”
Unlike competitors who often hide their reasoning processes, Mistral models show their complete thinking chain for users, and crucia, in the native language of the user instead of breaching English. “Here we have as the complete chain that is given to the user, but in their own language, so that they can read it real, see if it makes sense,” Lampra explained.
The company published two versions: Magistral Small, an open source model of 24 billion parameters, and masterful medium, a more powerful patented system available through the Mistral API.
Why Mistral AI models gained unexpected superimers
The models demonstrated surprising abilities that arose the training of duration. In general, the masterful environment retained multimodal reasoning skills: the ability to analyze images, even the thinking process focused only on mathematical problems and text -based coding.
“Something that we realized, not exactly by mistake, but something we absolutely expected, is that if at the end of the reinforcement learning training, plugs the initial vision, then splashing, you see Eagle.” “
The models also obtained sophisticated skills of functions calls, automatically conducting internet searches and multiple steps code execution to answer complex consultations. “What you will see is a model that does this, thinking, then realize, that’s fine, this information could be updated. Let me do a web search,” Lampra explained. “It will search the Internet, and then pass the real results, and will result in this regard, and will say, perhaps, perhaps the answer is not in these results. Let me search again.”
This behavior arose naturally without specific training. “It is something that white or not about the things to do next, but we discover that a real child of naturally is happening. It was a very pleasant surprise for us,” Lampra’s nutrition.
The advance of engineering that makes Mistral training faster than competitors
The Mistral technical team exceeded the important engineering challenges to create what Lampher describes as an advance in training infrastructure. The company developed a system for “online reinforcement learning” that allows AI models to continually improve answers, instead of depending on pre -existing training data.
The key innovation implied the synchronization of model updates in hundreds of graphics processing units (GPU) in real time. “What we did is that we found a way of unscrewing the model through GPU. I mean, from GPU to GPU,” Lampra explained. This allows the system to update the weights of the model in different GPU groups in seconds instead of the required hours.
“There is no open source infrastructure that will do this with this,” Lampra said. “In general, there are many open source attempts to do this, but it is extremely slow. Here, we focus a lot on efficiency.”
The training process was much faster and more cheaper than traditional pretending. “It was much cheaper than regular training. Previous training is something that would take week or months in other GPUs. Here, we are not close to this. It was like, I put it in how many people we put on this.
Nvidia compromises 18,000 chips on European independence of AI
The Mistral Compute platform will be executed in 18,000 or the new Grace Blackwell chips in Nvidia, initial lodged in a data center in Essonne, France, with expansion plans in Europe. The CEO of Nvidia, Jensen Huang, described the association as a crucial for European technological independence.
“Each country should build AI for its own nation, in its nation,” Huang said in a joint announcement in Paris. “With the Mistral AI, we are developing AI models and factories that Servereign platforms for companies from all over the intelligence in all industries.”
Huang projected that the computer capacity of Europe would increase ten times in the next two years, with more than 20 “AI factories” planned throughout the continent. Several of these facilities will have more than a capacity gigavatio, potentially classifying among the world’s largest data centers.
The association extends beyond the infrastructure to include NVIDIA’s work with other European AI and perplexity companies, the search company, to develop reasoning models in several European languages where limited training dates.
How Mistral plans to solve the environmental and remote problems of AIS
Mistral Compute addresses two main concerns about the development of AI: the environmental impact and sovereignty of the data. The platform ensures that European clients can maintain their information within the EU borders and under European jurisdiction.
The company has been associated with the National Agency for Ecological Transition of France and Carbone 4, a leading climate consultancy, to evaluate and minimize the carbon footprint of its AI models through its life cycle. Mistral plans to feed your data centers with decarbonized energy sources.
“When choosing Europe for the location of our sites, we give ourselves the ability to benefit from large -darcated energy sources,” the company said in its announcement.
Speed Advantage gives Mistral reasoning models to the practical shore
The early tests suggest that Mistral reasoning models offer competitive performance while addressing a common criticism of existing systems: speed. The current OpenAI reasoning models and others may take minutes to respond to complex consultations, limiting their practical utility.
“One of the things that people generally do not like this reasoning model is that he only thought he is intelligent, sometimes he is taking a long time,” Lampra said. “Here you really see the result in just a few seconds, sometimes less than five seconds, sometimes even less than this. And the experience changes.”
The speed advantage could be crucial for business adoption, where the waiting minutes for AI responses create workflow bottlenecks.
What the Mistral Infrastructure Bet means for the global competition of AI
The infrastructure Mistral movement puts it in direct competition with the technological giants that have dominated the cloud computing market. Amazon Web Services, Microsoft Azure and Google Cloud currently control most of the cloud infrastructure worldwide, while Diet’s players such as Coreweave have gained ground specifically in AI’s workloads.
The company’s approach differs from competitors by offering a complete vertical integrated solution, from hardware infrastructure to AI models and software services. This includes Mistral AI Studio for developers, Le Chat for business productivity and Mistral Code for programming assistance.
Industry analysts see Mistral’s strategy as part of a broader trend towards the regional development of AI. “Europe urgently needs to expand its AI infrastructure to stay competitive worldwide,” Huang observed, echoing those who express themselves with the voice of European politicians.
The announcement occurs when European governments are increasingly concerned about their dependence on US technology companies for the critical infrastructure of AI. The European Union has pledged 20 billion euros to build “gigafactories” of AI throughout the continent, and the association of Mistral with Nvidia could help accelerate those plans.
The dual announcement of Mistral of infrastructure and model capacities indicates the company’s ambition to become an integral artificial intelligence platform instead of just another model supplier. With the support of Microsoft and other investors, the company has raised more than $ 1 billion and continues to seek additional funds to support its expanded scope.
But Lample sees equally greater possibilities for reasoning models. “I think that when I look at progress internally, and I think that at some reference points, the model had more than 5% accuracy every week to, perhaps, like six weeks in total,” he said. “So it is improving very fast, there are many, many, I mean, tons of ideas, you know, small ideas in which you occur to improve performance.”
The success of this European challenge to the domain of the American AI may ultimately depend on whether customers value sovereignty and sustainability sufficient to change established suppliers. For now, at least, they have an option.
