French startup Mistral AI is attacking OpenAI’s dominance well under a year after its creation. After cutting its teeth on small artificial intelligence models and raising nearly 500 million euros, the young company introduced Mistral Large this Monday. This very large language model aims to compete with OpenAI’s GPT-4 and Gemini from Google, the world’s most powerful AI models.
To achieve its goals, Mistral has decided to fight fire with fire, because, like OpenAI, it has signed an exclusive distribution partnership with Microsoft for its new flagship model. The agreement is accompanied by a minority investment in the company’s capital (the exact amount of which is kept secret) from the American giant, which has invested more than 13 billion in OpenAI. While Mistral presented itself as a champion of open source during its operational launch in September 2023 and signaled an alternative path, it now offers its best models under the same access mode as OpenAI. Was this the price to pay to compete with the best?
Mistral leaves open source for performance
At its launch, the Mistral presented itself as a new. Spear » Open models, a role that Meta previously assumed with its range of Llama models. ” We believe that generative AI requires an open approach
», the company wrote, Technology to fight censorship and bias that will shape our future. » Open source (a publication with free access to numerous technical information on the model, editor’s note) was presented as the ideal way to deal with it. The emerging oligopoly “, i.e. establishing a market composed of only a few American players, including the OpenAI-Microsoft duo and Google, followed by a few very well-funded startups (Anthropic, Cohere, Inflection AI…). Only Meta took this path. .Since the launch of its first model in September 2023, the startup has attracted AI experts through its approach and performance compared to closed models. A second model, released three months later and nicknamed “Mixtral” by its users, finally convinced the sector of the seriousness and potential of young shoots. The latter suggested the possibility that the open source champion could surpass the performance of the American giants. But from November 2024, Arthur Manns explained that the Mistral models will not all be open. During the conference Artificial Intelligence Marseille Organized by GalleryHe has already explained that the fact of initially opting for the release of open source models, “That doesn’t mean we don’t develop proprietary models”.
Henceforth, the startup separates its models on its site “ open “, and its models” optimize » Synonym for off. In other words, its activity includes two categories. Arthur Mance fully assumes this strategy in his interview the world, and makes it clear that it was planned from the start, in addition to vouching for his company’s future investment in open source. It remains Mistral, who wanted ” Close the performance gap between black box and open solutions » At its launch, it finally bridged the gap into its own solutions. At X, company co-founder and scientific director Guillaume Lampel The new Mistral presents a small closed model
which is only accessible through the startup’s platform or through a Microsoft cloud such as Mistral Large, such as ” A significantly more efficient option » than its open source counterpart Mixtral…Mistral on the Imperial Route
By moving away from open source, Mistral partially solves the problem of monetizing its models. Certainly, it has signed agreements with the biggest cloud hosts (Microsoft Azure, Google Cloud and Amazon Web Services) to distribute its models for remuneration, as well as launched its own platform to distribute them, aptly named La Platform. But at the same time, thanks to open source, its models can be used by any company without going through Mistral or Host. By switching to closed mode, users will be forced to reach into their wallets to use the Mistral Large or Mistral Small models, whether directly through Mistral or through the Microsoft Cloud.
Companies will inevitably be tempted, because with Mistral Large, the startup finally has its flagship product, demonstrating that it can really do as well. In her model presentation blog, she uses no less than seven Benchmark (Performance Tests, Editor’s Note) To prove his point, the resulting graph below. It performs better than Google and Anthropic’s AIs, only being beaten overall by GPT-4, while being better at specific tasks.
To further its efforts to attract companies, Mistral has released an equivalent of ChatGPT called Le Chat, so they can test the capabilities of Mistral Large, which serves as the tool’s engine.
Not only does the Mistral do it better, but it does it with less. Mu The Wall Street Journal, Arthur Mensch explains that Mistral Large cost less than 20 million euros to produce, where OpenAI director Sam Altman explained that it cost between 50 million and 100 million dollars to train GPT-4, more than double that. The French startup also has much less funding than its competitors, and its team size is much smaller. This has been the startup’s mantra since its inception and so far it has been delivering on its promises.