Intern blunder? Hacker? On December 8, Mistral AI’s profile on the social network a series of mysterious characters. Enigmatic for laymen only. AI pros know perfectly well what’s going on: the French start-up has just published a link to its new major language model: Mixtral 8X7B. No fuss, no cleverly staged demo. Raw info for purists. “Just tweeting download links without any context is incredibly cool,” laughs one community member on X.
But the French AI child prodigy is not only attracting developers: the company has at the same time completed a spectacular fundraising (385 million euros), led by the Americans Andreessen Horowitz and Lightspeed Venture. “For a French start-up that is only seven months old, it seemed impossible,” agrees, with a touch of admiration, a Parisian tech entrepreneur.
By making the two announcements coincide, Mistral AI succeeded perfectly. The major European regulation of AI (AI Act) has just been validated by the Brussels trilogue (Council, Commission, Parliament), and NeurIPS, the great rally of artificial intelligence researchers, was held throughout the week (December 10 to 16). All the spotlight is therefore on the sector. Laudatory comments from professionals flourish alongside those from politicians and media articles. “It really is the French genius as we like to see and celebrate it,” greeted the President of the Republic Emmanuel Macron, during a trip at the start of the week. “The performance and lightness of the Mistral models are impressive,” confides Boris Lecoeur, CEO of Cloudflare France, which has just added a Mistral model to its Workers AI platform.
“Foundation models are the most courageous level”
A real strike. Which should not, however, make us forget the complex challenges on the road to the new French unicorn. The first ? Find an economic model adapted to your nature open source. The choice to make the structure of these models public is interesting for several reasons. L’open source allows innovators to move forward faster by pooling their advances, and allows civil society to have better control over the technology developed and the risks it may pose. Without theopen source, the Internet would also be much less sophisticated than it is today. It is in this community that, for example, the idea of integrating the audio and video formats omnipresent today into Wb browsers emerged.
For businesses, theopen source however, poses an economic challenge: how can you make your research and work profitable if competitors can easily re-exploit it? Over the last fifteen years in software, companies have invented new models (Red Hat, for example, values the support provided by its teams on its products). In AI, the sums initially invested to create foundation models are much more substantial: training these models costs tens of millions of euros.
The OpenAI leader who initially had an approach open source, has also changed its tune and now keeps its developments secret. Mistral AI currently gives few details on how it intends to generate revenue. “They have a very “Silicon Valley” approach: raise a lot of money in order to grow very quickly, create a community and refine their business model“, analyzes a professional in the sector.
Mistral AI appears to be moving towards a hybrid path with open base models and more powerful (or perhaps more customized) closed iterations. However, it remains to stand out from the competition. The start-up has positioned itself in the most technical segment of generative AI: that of foundation models which, as their name suggests, are the first layer of the building. Next comes the operational layer (connectors between databases and tools) then the application layer (image editing tool, etc.). Few companies have made the same choice as Mistral AI.
“Foundation models are the bravest and hardest level,” analyzes Vincent Luciani, CEO of Artefact, a consulting firm specializing in AI. And the few players who have launched into it generally enjoy significantly greater financial resources: DeepMind belongs to Google, Anthropic obtained 4 billion dollars from Amazon and 2 billion from Google for good measure. And of course, the leader OpenAI, whose latest model is at the top of the comparisons, has a close partnership with Microsoft which has invested 10 billion dollars in it.
Thierry Breton’s tackle
Third unknown hovering over Mistral AI’s head: the impact of the AI Act, the new European regulation on artificial intelligence. Thierry Breton, the European Commissioner for the Internal Market, made it clear that the demands of companies in the sector like Mistral AI were not his priority. “Mistral AI is lobbying, that’s normal. But we are not fooled by anything. It is defending its business today, and not the general interest,” he declared to There Tribune on November 24.
Under the leadership of France, Germany and Italy, worried about the health of their national gems of artificial intelligence, the EU has, however, watered its wine. The technical transpositions of the orientations on which the European trilogue agreed on December 8 must still be determined in the coming weeks. Details that will have a significant impact on European AI start-ups.
To meet its challenges, Mistral AI nevertheless has many assets in its game. First, strong political support from France. President Emmanuel Macron included Arthur Mensch, co-founder of Mistral AI, in an AI Gen committee tasked with advising the government on French AI strategy. And he actively advocates for regulation of technology conducive to the development of start-ups. The relationships that Mistral AI will build with Salesforce will also be worth watching. The American, who participated in the latest fundraising of $385 million, in fact has an empire in business software (customer relationship management, etc.). This could be a powerful accelerator for the distribution of Mistral technologies.
The team behind Mistral AI has, finally, an extraordinary background. Arthur Mensch, the CEO, went to Polytechnique and Normale sup before working for Google DeepMind. The scientific director of Mistral, Guillaume Lample (also X2011) is, for his part, one of the creators of the LLama language model of the giant Meta. A graduate of ENS, Timothée Lacroix, the technical director, also worked for eight years in the Parisian AI laboratory of Facebook’s parent company. “They have academic training that is among the best in the world and have been able to train on the most powerful infrastructures,” summarizes Vincent Luciani.
The secret recipe of Mistral AI
If Arthur Mensch was the first to confide to the newspaper The echoes that six months ago”[il] didn’t even know what an investor was”, the trio was also guided by experts on these subjects: Cédric O, former Secretary of State for Digital from 2019 to 2022 as well as Jean-Charles Samuelian and Charles Gorintin, respectively CEO and technical director of the start-up Alan, one of the French unicorns. “Technical profiles are behind the greatest technological successes – Mark Zuckerberg with Facebook, Steve Jobs with Apple, argues Cédric O. This is why funds such as Andreessen Horowitz are so popular.”
If Mistral AI has managed to build such high-performance models in just a few months, it is largely thanks to the trio’s expertise in finding and selecting the best training data (which they keep secret). “Very few people know where to retrieve interesting data. Most use a corpus called Common Crawl, but there are many others less known. Also few people know how to organize training data optimally. is a true work of art,” explains the CEO of Artefact. It is indeed better to choose your “ingredients” carefully because training large language models is very expensive: it generally requires a “farm” of at least a thousand specialized graphics cards which cost between 20,000 and 40,000 euros each. If the mixture put in the oven is not the right one, you can quickly end up with a salty addition and a completely unsuccessful dish.
.