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

You have the mistralThe French artificial intelligence has declared a sweeping expansion to the European American cloud calculation giants, which responds to the American cloud computing giants of Europe, in response to American cloud computing giants on Wednesday, as well as Openai’s best models.

Paris-based company emerged Calculate the MistralA comprehensive AI infrastructure platform established as partners NvidiaDesigned to give an alternative to relying on cloud providers who like European businesses and governments Amazon Web Services, Microsoft Azureand Google Cloud. The action represents a significant strategic turn next to the Mistral to develop AI models to manage the whole technology stack.

“This movement remains a transformative step to the Mistral AI, because the Mistral is a transformative vertical speech for the AI,” said Arthur Mensch, CEO and Mistral AI co-founder. “This slip is responsible for ensuring that our solutions support only innovation and the EU adoption and support the technological autonomy of Europe.”

How Models of Thinking Thinking In No Language Build

In addition to the infrastructure announcement, the Mistral presented it Highway Reason Models Series – AI systems that are able to think of a step-like step-like step-by-step O1 model and China’s DeepSeek R1. However, Guillaume, the head of the Mistral, the company’s approach is different from competitors, he said.

“We wanted to learn everything from scratch, mainly, we wanted to learn our experience,” We wanted to learn our elasticity, “he said.

Unlike opponents who often hide their views, the Mistral models show users and very well-thought-out thinking in the native language of users in English language in English. “Here we are granted to the user, but we like the chain of thinking in their own languages, so they can read it, they will be able to read.

The company left two versions: HighwayOpen source model of a parameter of 24 billion Highway averageThere is a more powerful property system via the Mistral API.

The Mistral AI models won unexpected superpowers during training

The models demonstrated surprising opportunities arising during training. Basically, high-level medium, multimodal thinking – the ability to analyze images – even if the training process is only focused on text-based mathematical and coding problems.

“Something, mistakenly, something we waotely waited, in the end of the fact that we are definitely waiting for, suddenly, you suddenly think that nothing is considering the model,” he said.

Models also won the complex function call skills, automatically implementing multi-step Internet search and code execution to respond to complex queries. “It’s a model to do this, think, and then carry out this information, it will be updated.” It will search for a website. “

This behavior suddenly emerged without a natural training. “We saw that there was or not the next job, but it really happened in a natural way. Thus, it was a wonderful surprise for us.”

Engineering leapt that the Mistral coaches make faster than competitors

The technical team of the Mistral exceeded significant engineering problems to create the materials described as a progress in the training infrastructure of the material. The company has prepared a system for “online reinforcement learning”, which allows the AI ​​models to pre-existing training information, which allows answers to constantly improve.

The main innovation is synchronizing model updates between hundreds of real-time graphics processing units (GPU). “What we do is find a way to open a model via GPU. I mean, from GPU to GPU” Lapare explained. This allows the system to update the model weights between different GPU clusters, not usually the required hours of the system, but different GPU clusters.

“There is no open source infrastructure that will do it in the right way,” he said. “Typically, there are many things like open-source attempts to do this, but extremely slow. Here, we payed a lot of attention to effectiveness.”

The training process proved faster and cheaper than traditional exercise. “This was cheaper than the ordinary exercise. Here is something that will take a few weeks in another GPUs. Here, do not approach it.

NVIDIA operates 18,000 chips to Europe’s independence of the EU

This Calculate the Mistral The platform will work for the newest 18000 NVIDIA Grace Blackwell chipsInitially, in Essonne, France, located in a data center with expansion plans in Europe. Nvidia CEO Jensen Huang considered the partnership very important for European technological independence.

“Every country should build a AI in its nation, in its nation, in its nation,” Huang said in Paris. “Mistral AI, we develop models and AI factories that serve as sovereign platforms for enterprises in Europe to scalt intelligence among businesses.”

Huang predicted Europe’s AI computing for the next two years, and more than 20 “AI factories” planned. Several of these facilities will have more than a potential, which has a potential ranking between the world’s largest data centers.

Partnership, NVIDIA’s other European company and a state of confusion, searching company, expands to develop substantial models in various languages ​​where training information is often limited.

The Mistral plans to solve the EU’s environmental and sovereign problems

Calculate the Mistral Two major concerns about the development of AI, environmental affect and data sovereignty. The platform ensures that European customers can store their information in the EU boundaries and the European jurisdiction.

The company is partner with national crossing and carbon 4 to the leader climate consultant for environmental transition and carbon 4, leading climate consultant 4, AI models and 4-of carbon 4 to minimize. The Mistral plans to strengthen data centers with decarbonated energy sources.

“We are able to use Europe to accommodate Europe by selecting Europe, the ability to take advantage of decarbon energy sources,” he said.

Speed ​​advantage gives a practical edge of the Mistral Reasoning Models

Early test offers to convey the competitive performance while touching the general criticism of the existing systems of the Mistral’s reasoning models. Openai and other rationale models, which are based on practical assistance, may take a few minutes to respond to complex queries.

“Although one of the things that people dislike about this reason, sometimes it takes a lot of time,” he said. “Here you really see a few seconds, sometimes less than five seconds, sometimes less than that. And it changes the experience.”

The speed of speed can prove the importance of working for the waiting period waiting for the waiting minutes for the business setting.

The infrastructure of the Mistral means the Bet of the Mistral for the Global AI Competition

The transformation of the Mistral’s infrastructure puts it directly with the dominant technology giants in the cloud computing market. Amazon Web Services, Microsoft Azureand Google Cloud Currently, new players manage the majority of cloud infrastructure Coreweave They have gained space in AI workloads specifically.

The company’s approach is supported from the device infrastructure by offering a complete, vertical integrated solution from competitors – from the device infrastructure to the Software. This includes You have the Mistral Studio For the creators, Pussy Institution productivity and Mistral code for programming assistance.

Industrial analysts see the strategy of the Mistral as part of a broader trend towards the development of regional AI. “If Europe wants to be competitive in the world, it is necessary to scalt the AI ​​infrastructure,” Huang “was observed, there are concerns voiced by European politicians.

The announcement, because European governments, because European governments are becoming more and more concerned about the dependence of American technology companies for critical AI infrastructure. The European Union created 20 billion euros for the preparation of AI Gigafactories on the continent and Partnership with the Mistral Nvidia can help accelerate these plans.

Binary announcement of the Mistral infrastructure and model capacity, the company’s passion is not just another model provider, but to be a comprehensive AI platform. Support from Microsoft and other investors, the company has increased more than $ 1 billion and continues to look for additional financial to support the widened scope.

However, for example, sample-based models sees greater opportunities. “I think that when I look at the progress and some criteria, the model gained a 5% accuracy for a plus every week, perhaps six weeks.” “So it improves very fast, there are many, many, I want to say, tons of tons, you know, small ideas you know about it, will improve your performance.”

The success of this European problem is the success of the American AI dominance, which can eventually depend on the sovereignty and sustainability of customers. So far, at least there are options.