Enterprise leaders have joined a reliable program for nearly two decades. VB transform brings people to develop real enterprise AI strategies. Get more information
Incorrect AIThe launch of French artificial intelligence on Wednesday announced a clean expansion of AI infrastructure, which calls the company a response to Europe about the US Cloud Computing Giants, while simultaneously unveiling the new argument model that counters the latest Openic systems.
The company based in Paris revealed Incorrect calculationA comprehensive AI infrastructure platform built in a partnership, nvidiaDesigned to give European businesses and governments an alternative to relying on cloud providers in the United States Amazon Web ServicesFor, for, for,. Microsoft AzureAnd Google Cloud. This move represents an important strategic change for misunderstanding, from fullying the AI ​​model to control the entire technology stack.
“The move in the AI ​​infrastructure indicates a change move for the wrong AI, as it allows us to resolve a significant vertical vertical in the AI ​​value chain,” said Arthur Manishch, Mr. CCO and co -founder Arthur Manishch. “With this change, it is the responsibility to not only move our solutions to innovation and the adoption of AI, but also maintain the technical sovereignty of Europe and participate in its sustainability.”
How misunderstanding made models of reasoning that think in any language
In addition to the announcement of infrastructure, Mysterol unveils his magistrate series of reasoning models-the AI ​​system, like the open, is step-by-step logical thinking. O1 model And china DPSEC R1. But Mr. Galome Lampel, the chief scientist of Mr., says the company’s point of view is different from rivals in important ways.
“We did everything from the beginning, mainly because we wanted to learn our skills, like, what we do,” Limpel told me in an exclusive interview. “We really, on a strong online reinforcement pipeline, was really, very effective.”
Unlike rivals who often hide their reasoning process, Mr. Model’s models show their whole thinking to users – and importantly, instead of defaulting English in the user’s native language. Lamppal explained, “Here we have a fully thought that has been given to the user, but in their own language, so they can read it through it, see if it means it or not.”
The company released two versions: Magistral SmallA 24 billion parameter open source model, and Magistral MediumAnother powerful proprietary system available through Mr. API’s, is available.
Why Mr. AI models of Mr. received unexpected superpower during training
These models showed amazing abilities that emerged during training. Most importantly, the Magistral Medium maintained multilateral reasoning capabilities-the ability to analyze the pictures-though the training process is fully focused on text-based mathematics and coding issues.
Lamppal said, “What we felt, not at all, but what we did not expect at all is that if at the end of the short learning training, you plug an early vision encoder, then suddenly, you see the model to be able to argue on the pictures.”
These models also achieved sophisticated function calling capabilities to automatically find the multi -phase internet search and code for the answers to complex questions. Lamppal explained, “What you will see is a model, thinking, then feeling, okay, okay, this information can be updated. Let me like the web search.” “It will look like the Internet, and then it will actually approve the results, and the result will result, and it will say, maybe, the answer is not in the result. Let me find again.”
This behavior emerged naturally without specific training. Lamppal noted, “This is something that is on the next tasks or not, but we have come to know that it is in fact a natural way. So it was a great surprise to us.”
Engineering progress that makes Mr. Training of Mr
Mr. The technical team overcome the key challenges of engineering so that the lamps have described the progress in the training infrastructure. The company has developed a system for “Online Come Learning Learning” that allows the AI ​​model to permanently improve by creating answers instead of relying on the pre -existing training data.
Key innovation includes hundreds of graphics processing units (GPUs) to include model updates in real time. “What we have done is that we have found a way to open the model through the GPU,” said Lamppel. I mean, from GPU to GPU. ” This allows the system to update the weight of the model in various GPU clusters within seconds instead of the required hours.
The Lampel noted, “There is no one like an open source infrastructure that it will do properly.” “Generally, there are a lot of efforts like this open source, but it is very slow. Here, we focused a lot on performance.”
The training process proved to be much faster and cheaper than traditional pre -training. Lamppal said, “It was much cheaper than regular pre -training. Pre -training is something that will take weeks or months in the other GPU. Here, we are nowhere to it. It was like this, I rely on how many people we had put on it, but it was less.”
Nvidia committed 18,000 chips for European AI independence
Incorrect calculation The platform will run on the latest of Nvidia 18,000 Grace Black Wheel chipsIt is initially located in a data center in Esson, France, with expansion plans all over Europe. Nvidia CEO Jensen Huang described the partnership as important for European technical independence.
“Every country should make AI for its people in its own nation,” Huang said in a joint announcement in Paris. “With Mr. AI, we are developing models and AI factory that work as an independent platform for businesses across Europe to measure intelligence in industries.”
Huang predicted that in the next two years, Europe’s AI computing capacity will increase tenfold, with more than 20 “AI factories” planned throughout the continent. Many of these facilities will have more capacity than Gigwat, which potentially classifies the world’s largest data centers.
The partnership is beyond infrastructure, which can be included with other European AI companies of NVIDIA and to develop reasoning models in various European languages, search company, search company, where training data is often limited.
How misunderstanding AI plans to solve the environmental and sovereignty problems
Incorrect calculation AI solves two major concerns about development: environmental impact and data sovereignty. The platform ensures that European users can keep their information within the European Union borders and in the European jurisdiction.
The company has contributed with the National Agency and a leading climate consultation, Carbon 4, for environmental transfer, to evaluate and reduce the carbon footprint of its AI models throughout its lifetime. Misunderstandings have intended to strengthen their data centers through decorated energy sources.
The company said in its announcement, “By choosing Europe for the location of our sites, we provide ourselves with the ability to exploit widespread decorated energy sources.”
Speed ​​Advanced Mr. Provides Practical edge to Mustal’s reasoning models
Initial testing shows that Mr. Reception’s models offer competitive performance, while a joint criticism of existing systems – solving speed. The models of Openi and others may take minutes to answer complex questions, limiting their practical utility.
Lamppal noted, “One of those things that people do not usually like about this argument model is that although it is smart, it sometimes takes a lot of time.” “Here you really see output in just a few seconds, sometimes less than five seconds, sometimes less than that. And it changes the experience.”
The rapid advantage of adoption of a business can be important, where the minutes of AI’s answers create obstacles to the workflow.
What is the condition of Mr.’s infrastructure for the World AI competition
Mistress’s move in the infrastructure has directly competed with the technology giants that dominated the cloud computing market. Amazon Web ServicesFor, for, for,. Microsoft AzureAnd Google Cloud Currently control the majority of cloud infrastructure globally, while new players like Coriao Particularly the AI ​​has received the ground in the workload.
From hardware infrastructure to AI models, software services – the company’s perspective is different from competitors by offering a full, vertically integrated solution. This includes Mr.l AI Studio For developers, Chat For the productivity of the enterprise, and Invalid code To help programming.
Industry analysts see Mr. Mustal’s strategy as part of a wider trend toward regional AI development. Huang observed, “If he wants to be competitive globally, Europe needs to immediately increase its AI infrastructure.”
The announcement comes when European governments are rapidly concerned about their dependence on US technology companies for the AI’s major structure. The European Union has pledged $ 20 billion to build AI “Giga factories” throughout the continent, and Mr These projects can help accelerate.
The dual announcement of infrastructure and model capabilities indicates the company’s desire to become a comprehensive AI platform rather than the company’s only model provider. With the backing of Microsoft and other investors, the company has collected more than $ 1 billion and has continued to search for additional funding to support its expanded circle.
But the lamps look forward to the reason for the models of reasoning. “I think when I see the progress internalized internally, and I think, on some standards, the model is getting 5 percent of the accuracy every week, probably in six weeks,” he said. “So it is improving very rapidly, there are many, many, I mean, like tin, you know, small ideas that you can think about will improve performance.”
The success of this European challenge this European challenge can eventually depend on whether consumers value sovereignty and stability that can change with the providers. For now, at least, they have a choice.
