Mistral 7B, the first generative AI model from Mistral AI, the French “huge gamble”

Can Mistral AI be Europe’s answer to the domination of American companies over generative artificial intelligence models? Founded only at the beginning of 2023, the startup of three AI researchers has already raised 105 million euros in seed funding.

To realize its promises, however, Mistral AI must deliver solutions, and quickly. The young company is responding by announcing the availability to developers of its first open source model.

An Apache 2 licensed model with 7 billion parameters

“This first provision comes a little more than four months after the creation of the company, in accordance with its strategic roadmap”, emphasizes Mistral AI in a press release.

Called Mistral 7B, the LLM model is presented as “small size”. The ambition of the startup is in fact to design very specialized solutions that can be deployed on corporate infrastructures and are less expensive to operate.

Mistral 7B is described as small because it has 7 billion parameters, far from the hundreds of billions of general LLMs offered by Open AI. Its designers, however, assure that it “far surpasses the performance of the best model of 13 billion parameters available to date”.

At the same time, the new model makes it possible to halve the computing power necessary for its use (-46%). During an interview conducted before the announcement of Mistral AI, on September 19, Hanan Ouazan, Partner & Lead Generative AI for Artefact, confirmed the finalization of this first model.

Mistral AI in PoC in France

The expert from the firm specializing in AI highlighted the speed of execution of the startup’s researchers and the relevance of their choices. Artefact is collaborating with Mistral AI on a PoC for a client whose identity is not specified at this stage.

What sets the French startup apart is its focus on specialized and open source models. Offered under the Apache 2.0 license, the Mistral 7B code and its parameters can be reused and modified for commercial uses.

The LLM is “the first version of a series of models which will be gradually distributed over the coming months. It already allows its users to solve a large number of text processing or generation tasks.”

The model is accompanied by a deployment solution on existing cloud environments, Mistral further indicates. Finally, note that its teams are now working on “training significantly larger and more efficient models.”

A huge bet for France Digitale

To generate revenue, the company announces the rapid availability of “commercial solutions for hosting and model specialization for professional use cases.” The results of the European startup are followed very closely by decision-makers, politicians and investors in particular.

On the occasion of the Big Data & AI Paris 2023 exhibition, the general director of France Digitale, Maya Noël regretted the absence of a European champion of generative AI. “We need a much more mature tech ecosystem,” she reacted.

The investment of 105 million euros in Mistral AI aims to support this development. “It’s a huge gamble. Did we catch it on time? We hope so. We must try not to lag behind,” commented the CEO of France Digitale.

Otherwise ? “If we cannot develop our own LLM model, we must be very quick and responsive in our uses.” As such, she regrets the European desire to want to regulate, from now on, AI Gen through a “very very restrictive” framework.

This opinion is shared by several members of the new Generative AI committee set up by the French government to make proposals and adapt its national strategy on artificial intelligence.

Source link -97