Mistral AI, a highly anticipated French startup, has formally concluded its Series A funding round. The company has raised €385 million, or $415 million at the current exchange rate; Bloomberg estimates that these funds raise the company’s value to approximately $2 billion. Today, Mistral AI is also launching its commercial platform.
In order to establish a European competitor to OpenAI, Mistral AI secured a $112 million seed round less than six months ago. Co-founded by Google’s DeepMind and Meta alums, the company is constructing open technology-oriented foundational models.
Andreessen Horowitz (a16z) is overseeing the AI company’s most recent funding round, in which Lightspeed Venture Partners has also invested. In addition, an extensive list of investors, including Salesforce, BNP Paribas, CMA-CGM, General Catalyst, Elad Gil, and Conviction, are participating in the round.
Co-founder and CEO of Mistral AI Arthur Mensch said in a statement, “Since the inception of Mistral AI in May, we have been committed to a distinct course of action: to establish a European leader with a global presence in generative artificial intelligence, predicated on an open, accountable, and decentralized technological framework.”
Mistral 7B, the introductory model from Mistral AI, was introduced in September. Trained on a “small” dataset consisting of approximately 7 billion parameters, this large language model is not intended to directly compete with GPT-4 or Claude 2.
The company provided the Mistral 7B model as a free distribution rather than through application programming interfaces (APIs), allowing developers to utilize it on their personal devices and servers.
The Apache 2.0 license, which governs the model’s distribution, is an open-source agreement with no usage or reproduction restrictions other than attribution requirements. Although the model is accessible to all users, its development occurred in private, utilizing a proprietary dataset and undisclosed weights.
Mistral AI was also a significant contributor to the discourse surrounding the EU’s AI Act. The French AI startup has pushed for a complete exemption for foundational models, arguing that regulation should only apply to businesses and use cases involving goods that end users directly use.
Only a few days ago, EU legislators were able to reach political agreement. Organizations tasked with developing foundational models will be obligated to comply with transparency regulations by disclosing technical documentation and summaries of the datasets’ contents.
The top model from Mistral AI is currently only accessible via API.
The organization continues to have the intention of generating revenue from its fundamental models. As a result, Mistral AI is launching the beta version of its developer platform today. Through the use of APIs, other businesses will be able to pay to utilize Mistral AI’s models on this platform.
Developers will have access to the Mixtral 8x7B model (“Mistral-small”), in addition to the Mistral 7B model (“Mistral-tiny”). This model utilizes “a router network” to process input tokens and choose the best set of parameters before returning a result.
“This method controls cost and latency while increasing the number of parameters in a model, as only a fraction of the complete set of parameters is utilized per token.” Mixtral has a total of 45 billion parameters, but only 12 billion parameters are utilized per token. “Therefore, it generates output and processes input at the same rate and cost as a 12B model,” the organization stated in a blog post.
A free download of Mixtral 8x7B, which was also distributed under the Apache 2.0 license, is also possible. Mistral offers a Mistral-medium variant as its third option on its developer platform. It is reportedly more effective than the other models offered by Mistral AI, but it is only accessible via the paid API platform; a download link is not provided.
