Mistral AIThe French Artificial Intelligence Startup on Wednesday announced a comprehensive expansion in the AI ​​Infrastructure, which located the company as a reply to Europe for the American Cloud Computing Giants, as well as unveils new arguments models that rival the most advanced systems of Openai.

Paris -based company revealed Mistral ComputeA comprehensive AI infrastructure is created in partnership with platform NvidiaDesigned to give European enterprises and governments the option to rely on US-based cloud providers Amazon web services, Microsoft azureAnd Google CloudThis step represents an important strategic change for Mistral from developing a purely AI model to control the entire technology stack.

“This step in AI Infrastructure is a transformative step for Mistral AI, as it allows us to address an important vertical of the AI ​​price chain,” said Arthur Mench, CEO and co-founder of Mistral AI. “With this change, it is the responsibility of ensuring that our solutions not only run innovation and AI adoption, but also maintains the technical autonomy of Europe and contribute to its stability leadership.”

How has Mistral who thinks in any language made logic model

With the announcement of infrastructure, Mistral unveiled it Prime Series of Reasoning Model-AI Systems Step by-step Logical Thinking to Openaii O1 model And China Deepsek R1But Mistral’s chief scientist, Giloom Lampal says that the company’s approach is different from the contestants in important ways.

“We did everything with scratches, basically because we wanted to learn the expertise we have, like, in what we do,” Lampal told me in a special interview. “We really, like, managed to work on a very efficient online reinforcement learning pipeline.”

In contrast to contestants who often hide their logic processes, the model of Mistral displays the full range of ideas for users – and significantly, instead of defaulting in English in the original language of the user. “Here we like the full range of ideas that are given to the user, but in their language, so they can actually read through it, see if it makes sense,” Lample explained.

The company released two versions: Magistral smallA 24 billion parameter open-source model, and Magistral mediumA more powerful proprietary system is available through Mistral’s API.

Why Mistral’s AI model receives unexpected superpower during training

The model demonstrated the stunning abilities that emerged during the training. Most particularly, the magistral medium has maintained the ability to analyze the multimodal region capabilities-the training process was fully focused on text-based mathematical and coding problems.

“Something that we realized, not at all by mistake, but something we did not fully expect, is that if you plug the initial vision Encoder back at the end of the training of reinforcement, you suddenly, you suddenly, not out of anywhere, see the model being able to argue on images,” Lampal said.

The model also achieved sophisticated function-giving capabilities, automatically demonstrated multi-step internet discoveries and code execution to answer complex questions. “What you will see is a model who is doing this, thinking, then real, this information can be updated. I like a web search,” Lample explained. “It will discover like the Internet, and then it will actually pass the results, and it will result in it, and it will say, perhaps, perhaps the answer is not in this result. Let me find again.”

This behavior emerged naturally without specific training. “This is something that is on things to do further, but we found that it was really happening naturally. So it was a great surprise for us,” Lample said.

Pilgrimage

Mistral’s technical team faced important engineering challenges what Lample training describes as success in infrastructure. The company developed a system for “learning online reinforcement”, which allows the AI ​​model to continuously improve the reactions rather than relying on the training data already existing.

Major innovation included synchronizing the model updates in hundreds of graphics processing units (GPU) in real -time. “What we did is that we found a way to open the model through GPU. I mean, from GPU to GPU,” Lample explained. This allows the system to update model weight in various GPU groups within seconds instead of required hours.

“There is no one like the open source infrastructure that will do it properly,” Lample said. “Usually, there are lots of open source efforts to do this, but it is very slow. Here, we focused a lot on efficiency.”

The training process proved to be much faster and cheaper than the traditional pre-training. “This regularly was much cheaper than pre -training. Pre training is something that takes weeks or months on other GPUs.

Nvidia European AI gives 18,000 chips for freedom

Mistral Compute Platform will run on the latest 18,000 of NVIDIA Grace Blackwell ChipsInitially Essonne was placed in a data center in France, with a plan of expansion throughout Europe. Nvidia CEO Jensen Huang described the partnership for European technical freedom as important.

Huang said in a joint announcement in Paris, “Every country should build AI in its nation, for its nation.” “With Mistral AI, we are developing models and AI factories that serve as sovereign platforms for enterprises across Europe for intelligence information in industries.”

Huang speculated that Europe’s AI computing capacity would increase ten times over the next two years, with more than 20 “AI factories” plans throughout the continent. Many of these features will exceed a gigawatt of capacity, which is potentially ranking in the world’s largest data centers.

The partnership extends beyond infrastructure to include NVidia work with other European AI companies and search company, which is often limited to develop logic models in various European languages, where training data is often limited.

How to solve AI’s environment and sovereignty problems

Mistral Compute AI addresses two major concerns about development: environmental impact and data sovereignty. The platform ensures that European customers can keep their knowledge within the boundaries of the European Union and in the European jurisdiction.

The company has partnered with a major climate counseling, a major climate counseling of the French National Agency for Ecological Transition & Carbone, to assess and reduce the carbon footprint of its AI model during its life cycle. Mistral plan to provide strength to its data centers with decarbonized energy sources.

The company said in its announcement, “By selecting Europe for the location of our sites, we give ourselves the ability to benefit from decarbonated energy sources.”

Speed ​​Advantage gives a practical edge to the Reasoning Model of Mistral

Initial testing suggests that Mistral’s argument models provide competitive performance, addressing a common criticism of existing systems – speed. The current logic models of Openai and others can take minutes to answer complex questions, limit their practical utility.

“One of the things that people usually do not like this argument model, that is that even though it is smart, sometimes it takes a lot of time,” Lample said. “Here you really see output in a few seconds, sometimes less than five seconds, sometimes less than this. And it changes the experience.”

Speed ​​gains can prove to be important for adopting business, where workflow hurdles are formed for AI reactions.

What does Mistral’s infrastructure bets mean for global AI competition

Mistral’s move in the infrastructure has put it directly in competition with technology giants that dominate the cloud computing market. Amazon web services, Microsoft azureAnd Google Cloud Currently globally controls the majority of cloud infrastructure, while new players prefer Coreweave Especially received land in AI workload.

The company’s approach is different from the contestants by offering a full, vertical integrated solution – from hardware infrastructure to AI model software services. It also includes Mistral AI Studio For developers, Le chat For enterprise productivity, and Mistral Code For programming assistance.

Industry analysts look at Mistral’s strategy as part of a broader tendency towards regional AI development. Huang said, “Europe needs to increase its AI infrastructure immediately, if he wants to remain competitive globally,” Huang saw, resonating the concerns raised by the European policy makers.

This announcement comes when European governments worry about their dependence on American technology companies for important AI infrastructure. The European Union has committed € 20 billion for the manufacture of AI “Gigafacteries” in the continent, and Mistral’s partnership with Nvidia Can help accelerate those plans.

Mistral’s dual declaration of infrastructure and model abilities indicates the company’s ambition to become a comprehensive AI platform rather than just another model provider. With support from Microsoft and other investors, the company has raised over $ 1 billion and looking for additional funds to support its extended scope.

But lample argument also sees big possibilities for the model. “I think when I look at progress internal, and I think at some benchmarks, the model was getting 5% accuracy every week every week, such as probably, all, like six weeks,” he said. “So it is improving very fast, many, many, I mean, tons of tones, you know, small ideas that you may think that performance will improve.”

The success of this European challenge for American AI dominance may eventually depend on whether customers give importance to sovereignty and stability to switch from established providers. For now, at least, they have an option.