Bergur Thormundsson
Research expert covering technology in the Nordics, AI and smart cities
Mistral - statistics & facts
Mistral is Europe’s foremost offering in the ongoing generative AI competition. Coming relatively late to the stage in 2023, founded by former Meta and Google DeepMind AI researchers, it has already made a name for itself in the AI revolution. Through successive funding rounds the company has reached a valuation of just over six billion U.S. dollars, raising just over 500 million U.S. dollars in 2024. In a competitive market including juggernauts like Microsoft, Google, Meta, Amazon, and Apple, the funding is not as great as some U.S. AI startups, but the success of Mistral demonstrates that financials do not tell the entire story when developing new technologies.
Despite its size, MML2 is comparable to its larger competitors in capabilities. The model was barely four percentile points below GPT-4o, Llama 3.1, and Claude 3.5 Sonnet, with these three constituting the most advanced generative AI models available. Mistral’s models also perform exceedingly well in hallucination benchmarks, meaning they are far less likely than their competitors to give false or fabricated responses. Mistral’s models are all also trained in a large selection of languages, from English and French to Chinese and Hindi.
Open source is the way forward
Mistral’s unique offering over its competitors is two-fold. Firstly, many of its products, with the most recent ones being NeMo12b and MML2, are open-source and have been freely usable, though MML2 has some registration requirements. All the models predating MML2 were completely open to use. This contrasts with GPT-4o from OpenAI, or Gemini from Google, where their code base is locked, and usage requires a subscription. More of the larger AI makers, such as Meta, began advocating for open-source collaboration on AI in the summer of 2024, suggesting that Mistral was ahead of the curve in their large language model (LLM) development.Precise work and a focus on languages
The second unique aspect of Mistral's AI work is the size of its models. MML2 accounts for around 123 million parameters. That is impressive, though it pales in comparison to GPT-4 which is 14 times larger. The significance is that MML2 can be run on a small server with only a few graphic processing units (GPUs), whereas the larger models require entire server centers to run.Despite its size, MML2 is comparable to its larger competitors in capabilities. The model was barely four percentile points below GPT-4o, Llama 3.1, and Claude 3.5 Sonnet, with these three constituting the most advanced generative AI models available. Mistral’s models also perform exceedingly well in hallucination benchmarks, meaning they are far less likely than their competitors to give false or fabricated responses. Mistral’s models are all also trained in a large selection of languages, from English and French to Chinese and Hindi.
EU regulations a potential boon
In July 2024, both Meta and Apple announced their newest models would not be available in the EU due to regulatory disagreements. Willingly removing themselves from the 450 million strong user market of the EU means that Mistral now has fewer competitors in its primary market. It was founded and grown in Paris and is thus more accustomed to dealing with EU regulations than its American counterparts. Though OpenAI and Google have made no claims to leave the EU market, the lessened interest of major AI developers gives Mistral a stronger footing in Europe and a greater chance of growing even bigger in the coming years.
