A new addition has joined the Mistral AI herd. The French star of generative AI has already launched five major language models in twelve months. She just revealed one more, Mixtral 8X22B, presented on April 17 as being “cheaper, better and faster”. However, it is the rumors of a new fundraising in preparation, barely four months after the last, which agitates the AI landerneau the most. Contacted by L’Express, the French start-up did not comment. But according to the American media The InformationMistral AI is reportedly once again in discussions with investors to raise several hundred million dollars based on a valuation of $5 billion this time.
Signs that his “half-open door” policy is paying off. As critics like to point out, if Mistral AI defends an “open” approach, it is not completely open source. Understand by this that the company does not reveal all of its manufacturing secrets. Mistral has opted for a so-called hybrid approach open weights. It shares the parameters and architecture of many of its models, not the data on which they were trained.
But it’s probably the most sensible approach for a company in the sector. In the traditional software world, several players have shown that it is entirely possible to make money by sharing all of your developments. Red Hat sells, for example, the support provided by its teams on its products. But AI is a very different world. The costs of developing a large language model (LLM) are in no way comparable to those of traditional software. By the admission of OpenAI boss Sam Altman, the cost of training GPT-4 exceeds the threshold of 100 million dollars – and he did not specify whether it exceeds it by a little or very much .
The secret ingredient of Mistral AI
In this context, the approach open weights from Mistral AI is an interesting compromise. It allows anyone who wishes to use its language models and create new products from this base, while leaving the Frenchman with the secret ingredient that gives him a head start over many of his competitors: its expertise in data. “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 the training data optimally. is a real work of art”, explained to L’Express last December, Vincent Luciani, CEO of the AI consultancy Artefact, saluting the expertise of the trio of founders of Mistral at this level.
For Mistral, the challenge of sharing part of its advances also has obvious advantages. This allowed it to make a name for itself at a time when the world had its eyes on OpenAI. “It’s also a strong argument for attracting highly sought-after profiles,” confides a player in the sector. Experts capable of creating foundation models are not common in the streets, and are being chased away at high prices by tech giants. However, the possibility of publishing their work is often close to their hearts.
Certainly, Mistral does not open all versions of these language models and is moving towards a hybrid path with open LLMs and more powerful closed versions. But at a time when the majority of its competitors have opted for a completely closed approach (OpenAI, Anthropic, Aleph Alpha, Poolside, etc.), it continues to be an island of resistance. As long as the performance gap between its open models and its secret versions does not widen excessively, it will be able to legitimately continue to claim this chapel.
.