Mistral AI is the French startup that opted for efficiency before Deepseek. His future is uncertain

Mistral ai is the French technological jewel. The AI ​​startup has become practically the only European representative that competes with large companies and technological startups in the US or China. He also does it with an absolute focus on efficiency, which is just what is now valued as Deepseek. However, its future is complex.

A promising European startup. When we talk about her for the first time in 2023, he surprised that without having any product He managed to raise 105 million euros investment Soon, yes, the fruits of such commitment would begin to appear, which also generated that the firm raised More financing rounds.

Flag efficiency. One of the defining characteristics of Mistral AI was that his work was always oriented to do more with less. To seek efficiency. The startup, knowing not being able to compete with the Bigt Tech in resources, has always sought to create more compact models but with great performance. He succeeded with Large 2 123b, which was three times smaller that calls 3.1 405b but matched that model and others like Claude 3.5 or GPT-4o in some metrics.

They return to the load with Small 3 24b. The startup He has just announced The availability of Small 3 24b, its new “small” model. Actually it is not so much: those with size between 1b and 14. However, it is certainly an interesting llm for one thing: it competes from you to you with flame 3.3 70b – the last of the finish line, almost three sometimes bigger – and also does it with a latency (time it takes to appear every token) much smaller.

The performance is fantastic. Latency, if we pay attention to the internal tests of Mistral, fantastic.

Answer three times faster than the goal model. Its performance is comparable to GPT-4o Mini, and also exceeds QWEN-2.5 32B latency, which is something better in some benchmarks. This model, yes, has just had successor with QWen2.5-Max.

Open Source, and European data. Another advantages of Mistral’s AI is that it is an Open Source AI –“Open Weights”, rather-, as he calls, Deepseek or Qwen. Unlike them, Mistral is a startup that raises data governance in the EU, something that can be inert for government agencies and European companies. It has already been seen how there is suspicion about where the data ends From our chats with Deepseek –Italy is investigating In this regard – and this is an undoubtedly striking option.

Interesting uses. Mistral developers explain that their model is perfect for conversational assistants, because in them precisely matters that they respond quickly. They also highlight the ability to customize/polish the model to specialize in certain tasks such as legal council, medical diagnosis or technical support. The model is available on platforms such as Hugging Facebut it can also be executed at home in quantized seeing if you have at least 4090 or for example a MacBook with 32 GB of memory. At the moment it does not seem available In le chatthe web service that has confessed to us to be based on Mistral Large 2.1.

Reasoning model in sight. But while their competitors are launched to the race for the reasoning models, Mistral has left something behind in this area. In a message in x They clarify that Small 3 does not use synthetic data “which makes it a great base for anything in reason.” In their official announcement they go further and point out that “among many other things, in the coming weeks, large and small Mistral models are expected with greater reasoning capacity.”

Deepseek advancing on the right. The Chinese fashion startup has become the star of the moment with its models, and especially with its reasoning model, Deepseek R1. He has also achieved it using the same weapon as Mistral: efficiency. Deepseek’s success validates Mistral’s strategy, of course, but the question is whether Mistral will continue to compete or be overwhelmed by Chinese and US startups if they also partially or totally adopt that same approach.

Mistral’s market share is modest, but is in danger of reducing with the strong competition of US and China companies. Source: FT.

Doubts about the future. In media as Financial Times The concern about the future of Mistral and the European startups is discussed. There is no other to work in LLM –The German Aleph Alpha He left that race In September 2024-, and that compromises the future of European efforts to compete and not depend completely on the models of China and the US no matter how they can be. In Spain projects such as Alia are an interesting first step, but for now they are far from the LLMs and models mentioned.

Operational limbo. This economic newspaper also indicates that with an assessment of 6,000 million euros Mistral is in a kind of limbo. He has raised too much money to gradually disappear, but not enough to compete with US mega -companies, for example.

Acquihire on the horizon? Sean Maher, from the Entext consultancy, believes that Mistral can Follow the steps of inflection aiwhich was acquired by Microsoft – Mustafa Suleyman, current head of Ia in Redmond, was his co -founder. Thus, a potential acquisition of its resources and especially of their talent (which is usually called ‘acquihire’) is not ruled out.

Image | Mistral ai