French startup Mistral AI has officially closed its much anticipated Series A funding round. The company has raised €385 million, or $415 million at today’s exchange rate — according to Bloomberg, it values the company at roughly $2 billion. Mistral AI is also opening up its commercial platform today.
As a reminder, Mistral AI raised a $112 million seed round less than six months ago to set up a European rival to OpenAI. Co-founded by Google’s DeepMind and Meta alums, Mistral AI is working on foundational models with an open technology angle.
Andreessen Horowitz (a16z) is leading the most recent funding round, with Lightspeed Venture Partners investing once again in the AI company. That’s not all, as a long list of investors is also participating the round, such as Salesforce, BNP Paribas, CMA-CGM, General Catalyst, Elad Gil and Conviction.
“Since the creation of Mistral AI in May, we have been pursuing a clear trajectory: that of creating a European champion with a global vocation in generative artificial intelligence, based on an open, responsible and decentralised approach to technology,” Mistral AI co-founder and CEO Arthur Mensch said in a statement.
In September, Mistral AI released its first model called Mistral 7B. This large language model isn’t meant to compete directly with GPT-4 or Claude 2 as it was trained on a “small” dataset of around 7 billion parameters.
Instead of opening access to the Mistral 7B model via APIs, the company made it available as a free download so developers could run it on their devices and servers.
The model was released under the Apache 2.0 license, an open source license that has no restrictions on use or reproduction beyond attribution. While the model can be run by anyone, it was developed behind closed doors with a proprietary and undisclosed dataset.
Mistral AI also played an important role in shaping the discussions around the EU’s AI Act. The French AI startup has been lobbying for a total exemption for foundational models, saying that regulation should apply to use cases and companies working on products that are used by end users directly.
EU lawmakers reached a political deal just a couple of days ago. Companies working on foundational models will face some transparency requirements and will have to share technical documentation and summaries of what’s in the datasets.
Mistral AI’s best model is now only accessible via an API
The company still plans to make money from its foundational models. That’s why Mistral AI is opening its developer platform in beta today. With this platform, other companies will be able to pay to use Mistral AI’s models via APIs.
In addition to the Mistral 7B model (“Mistral-tiny”), developers will be able to access the new Mixtral 8x7B model (“Mistral-small”). This model uses “a router network” to process input tokens and choose the most apt group of parameters to give an answer.
“This technique increases the number of parameters of a model while controlling cost and latency, as the model only uses a fraction of the total set of parameters per token. Concretely, Mixtral has 45B total parameters but only uses 12B parameters per token. It, therefore, processes input and generates output at the same speed and for the same cost as a 12B model,” the company wrote in a blog post.
Mixtral 8x7B has also been released under the Apache 2.0 license and is available as a free download. A third model, Mistral-medium, is available on Mistral’s developer platform. It supposedly performs better than Mistral AI’s other models and it is only available through the paid API platform — no download link available.
Comment