Mistral AI, the French startup once dismissed as Europe's underfunded answer to OpenAI, just opened public access to Large 4 — a model built with a trillion parameters, nicknamed internally "le Chonk". It's the largest AI system ever trained end-to-end on European soil, with no American or Chinese cloud involved.
A trillion parameters doesn't mean a trillion calculations per question, though. Large 4 runs on a mixture-of-experts design: the model keeps a huge roster of specialized sub-networks in memory, but each query only activates around 49 billion of them — whichever specialists the task actually needs. Think of a building with hundreds of offices where the lights only switch on in the rooms being used.
Training took roughly two months from scratch on about 4,000 Nvidia Grace Blackwell chips, all housed in Mistral's own data centers in France. The dataset spans more than 160 languages, including every official EU language, and the model handles a context window of up to a million tokens.
On Mistral's own benchmarks, Large 4 scores well on cybersecurity tasks — 93% on the Cybench security exercises, for instance — and holds up in coding and financial-analysis tests. None of that has been independently verified yet, and rivals like Alibaba, DeepSeek and Zhipu test against different setups, so direct comparisons are shaky at best.
The launch matters for Europe specifically. Open-weight AI has so far been a contest between American and Chinese labs — Meta, DeepSeek, Alibaba, Moonshot. Mistral has leaned on French government backing and investor money under the banner of "digital sovereignty" for years; Large 4 is the first real proof that bet might be paying off.
Full weights are promised for October 27. Only then will outside developers get to run Large 4 on their own hardware and test Mistral's numbers for themselves.
