Microsoft-backed Mistral launches European AI cloud to compete with AWS and Azure

Mistral AI announced the start of French artificial intelligence, on Wednesday, a comprehensive expansion of the Amnesty International infrastructure that puts the company as an answer in Europe to the American cloud computing giants, while simultaneously reveals the new thinking models that compete with the most advanced Openai systems.

The Paris -based company has revealed the poor security, which is a comprehensive platform for Amnesty International’s infrastructure that was built in partnership with NVIDIA, designed to grant European institutions and governments an alternative to relying on US cloud service providers such as Amazon Web Services, Microsoft Azure and Google Cloud. This step represents a major strategic shift for purely developing models to control the entire stack of technology.

“This transition to Amnesty International’s infrastructure represents a transformative step to treat wrong artificial intelligence, as it allows us to have a critical vertical treatment of the series of the value of artificial intelligence,” said Arthur Minche, CEO and co -founder of Mistral Ai. “With this transformation, the responsibility of ensuring that our solutions are not only creating and adopting artificial intelligence, but also adhere to technological independence in Europe and contributing to its sustainability.”

How was the wrong thinking models that think about any language

Besides the infrastructure announcement, Mistral revealed a series of models of judicial logic-artificial intelligence systems capable of logical thinking by step similar to the Chinese OPEN and Deepseek R1 model. But Guillaume Lample, the chief scientist in Mistral, says that the company’s approach is different from competitors in decisive ways.

“We did everything from scratch, mainly because we wanted to learn the experience that we have, such as the flexibility in what we do,” Lample told me in an exclusive interview. “We have already been able to be, like, very effective in the strongest online learning pipeline.”

Unlike competitors who often hide their thinking processes, Mistral models display a series of full ideas for users – decisively, in the mother language of the user instead of virtual English. “Here we have a series of full thought that is given to the user, but their own language, so that they can actually read it, and know if it is logical,” I explained Lample.

The company has released two versions: Magistral Small, an open source model 24 billion, and in the middle of the judiciary, which is a more powerful system available through the Mistral Application interface.

Why did artificial intelligence models in Mistral have gained unexpected great powers during training

Models showed sudden abilities that appeared during training. It is worth noting that the average judicial retains the capabilities of multi-media thinking-the ability to analyze images-although the training process focused only on mathematical problems and text-based coding.

“Something we realized, not completely by mistake, but something we never expected, is that if at the end of the reinforcement learning training, you will deliver the first vision encrypted, then suddenly, somewhat of nothingness, see that the model is able to do the thinking of the images,” said Lample.

Models have also gained advanced capabilities to summon jobs, automatically performing multi -step searches and carrying out code to answer complex queries. “What you will see is a model that does this, thinking, then realizing, well, this information may be updated. Let me do like searching on the web,” explained Lamples. “You will search for the Internet like the Internet, and then the results will already pass, and it will result, and it will say, perhaps, the answer is not in these results. Let me search again.”

This behavior appeared naturally without specific training. “It is something whether or not in things to do next, but we found that it was already happening normally. So it was a very nice surprise for us,” Lamples pointed out.

Hate engineering that makes Mestra training faster than competitors

Mistral technical team overcame major engineering challenges to create what Lample describes as a breakthrough in infrastructure training. The company has developed a system for “learning online reinforcement” that allows artificial intelligence models to improve continuously while generating responses, rather than relying on pre -existing training data.

The main innovation ensures synchronization of models updates across hundreds of graphics processing units (GPU) in actual time. “What we did is that we found a way to decipher the model through graphics processing units. I mean, from GPU to GPU,” the clearest lamble. This allows the system to update the model weights across different groups of GPU within seconds instead of the hours usually required.

“There is no such open source infrastructure will do it correctly,” noted Lampled. “Usually, there are many open sources attempts to do this, but it is very slow. Here, we focused a lot on efficiency.”

The training process has proven faster and much cheaper than traditional training. “It was much cheaper than usual training before training. Before training is something that takes weeks or months on other graphics processing units. Here, we are not anywhere close to this. It was so, I rely on the number of people we put on this. But it was more like, it was less than one week,” said Lample. “

NVIDIA is committed to 18,000 chips to the independence of European artificial intelligence

The Mistral account platform will be operated on 18,000 of the latest Grace Blackweell chips in NVIDIA, which is initially found at the Essonne data center, France, with plans to expand all over Europe. The CEO of Nvidia Jensen Huang described the partnership as decisive to European technological independence.

“Every country must build artificial intelligence for their nation, in their nation,” Huang said in a joint declaration in Paris. “With Mistral Ai, we develop the models and factories of Amnesty International that act as sovereign platforms for institutions throughout Europe to expand intelligence through industries.”

Huang expected that the ability of computing in Europe to increase artificial intelligence over the next two years, with more than 20 “Amnesty International Factory” throughout the continent. Many of these facilities will have more than just capacity, and possibly arrange them among the largest databases in the world.

The partnership extends beyond the infrastructure to include NVIDIA’s work with other European artificial intelligence companies, the search company, to develop thinking models in different European languages ​​where training data is limited.

How Mistral plans to solve environmental and sovereignty intelligence problems

Compute Mistral addresses two main attention on artificial intelligence development: environmental impact and database. The platform guarantees that European customers can maintain their information within the borders of the European Union and under the European Judicial Wilayat.

The company has a partnership with the National Agency for Environmental Transition and Karbon 4, a leading climate consultant, to evaluate and reduce carbon mark in artificial intelligence models throughout their life cycle. Mistral plans to operate their data centers with non -carbon power sources.

“By choosing Europe to our website, we give ourselves the ability to benefit from the energy sources that the carbon has largely canceled,” the company stated in its announcement.

Speed ​​feature gives thinking models in a practical advantage

Early test indicates that Mistral thinking models provide competitive performance with the treatment of joint criticism of current systems – speed. The current thinking models of Openai and others can take minutes to respond to complex queries, which limits their practical benefit.

“One of the things that people usually do not like in this thinking model is that although it is smart, it sometimes takes a lot of time,” he pointed out. “Here you really see the output in just a few seconds, sometimes less than five seconds, and sometimes less than this. It changes the experience.”

The speed feature may be it is very important to adopt business, as waiting for the artificial intelligence responses creates workflow.

What Mistral’s Military Infrastructure Betthade means artificial intelligence

Mistral’s transition to infrastructure puts it in direct competition with technology giants who dominated the cloud computing market. Amazon Web Services, Microsoft Azure and Google Cloud services are currently controlling most of the world’s cloud infrastructure, while the latest players such as Coreave gained special reasons for the burdens of artificial intelligence.

The company’s approach to competitors is different by providing a complete and vertical solution – from devices infrastructure to artificial intelligence models to software services. This includes the Mistral AI Studio for developers, chatting for the productivity of the institution, and the Mistral code to help programming.

Industry analysts sees Mistral strategy as part of a wider direction towards developing regional artificial intelligence. “Europe urgently needs to expand the scope of Amnesty International’s infrastructure if it wants to remain competitive worldwide,” Huang noted the concerns expressed by European politicians.

This advertisement comes at a time when European governments are increasingly concerned about their dependence on US Guardian Infrastructure companies. The European Union has a value of 20 billion euros to build “Gigafactories” from artificial intelligence throughout the continent, and Mestral’s partnership with NVIDIA can help accelerate these plans.

Mistral’s double announcement indicates the infrastructure and typical capabilities that the company’s ambition becomes a comprehensive AI platform instead of just an other model provider. With support from Microsoft and other investors, the company has gathered more than one billion dollars and continues an additional financing request to support its expanded scope.

But Lample sees greater possibilities in front of thinking models. He said: “I think that when I look at progress internally, and I think it is in some criteria, the model was getting an extra accuracy of 5 % every week, perhaps, like six weeks in everything.” “Therefore, it is improving very quickly, there are many, many, I mean, many tons of such as, as you know, the small ideas you can think of and that will improve performance.”

The success of this European challenge to the dominance of American artificial intelligence may depend on whether customers appreciate sovereignty and sustainability enough to move from well -known service providers. At the present time, at least, they have a choice.