Technology

France Built a Big New AI You Can Run Yourself

Martin HollowayPublished 2month ago3 min readBased on 4 sources
Reading level
France Built a Big New AI You Can Run Yourself
Image by cookieone from Pixabay

Mistral AI has released Mistral Large 4, a large AI system that can work with both text and images and will be shared as open weights.

It uses a granular Mixture-of-Experts design. Think of it as a large team of specialists where only a few step in to answer each question.

The size is stated directly. The full system holds 1.05 trillion settings, but only 49 billion are used at a time. A separate 1.6 billion part handles images Mistral documentation. Its nickname is Le Chonk.

The rollout is staged. It is available now through an API and can be run from Europe on Mistral Cloud. Mistral AI says it was built end-to-end in Europe. The open-weights release is planned for the end of October announcement.

On results, Mistral AI describes Large 4 as the best open-weights model from the US or Europe on combined test scores. The company claims top results for cyber defense, manufacturing and finance, and says the model beats closed frontier models on visual grounding, which means linking words to the right parts of an image announcement. Mistral AI is working privately with cybersecurity partners on the model.

The chief executive put the security claim more bluntly. France's Mistral CEO said its new AI model beats Chinese models on cybersecurity Reuters. That comment came after France's Mistral raised €3 billion at a value of around €21 billion, about $24 billion Reuters.

The broader context for teams that run AI in daily work is access, location, and fit. An API now with open weights promised later gives a quick way to start while keeping the choice to run it yourself or adapt it later. Building and running it in Europe helps with buying rules about where data and AI work stay.

In my view, the design choice is the part to watch. Keeping many specialists stored away but using only a few each time can save work. For cyber defense, finance and factory jobs, where the language is narrow but mistakes are costly, that can help if questions still reach the right specialists when data changes. A lead in visual grounding, if outside tests confirm it, would matter more for paperwork-heavy and operations-heavy jobs than for everyday chat.

What this could make possible, if the schedule holds, is another top-level open-weight choice hosted in Europe. That gives more options for hospitals, banks and other regulated users without forcing a trade between power and control.