Mistral AI's trillion-parameter yardstick

Crossing the trillion-parameter mark is no longer the preserve of American labs or Chinese giants. With the preview of Mistral Large 4, nicknamed "Le Chonk", the French company is launching its largest model yet: a multimodal Mixture-of-Experts architecture with 1.05 trillion total parameters. According to Mistral AI's official documentation, the model activates 52 billion of them per token (although some outlets report 49 billion) and relies on a 1.6-billion-parameter vision encoder, with a context window of one million tokens. It is the first European architecture of this scale to be presented with a promise of open weights. It is also the first model released since the €3 billion funding round Mistral closed in September, as The Next Web reports.

For now, though, access is only available through the public preview API on Mistral Studio, which runs in the company's European data centers. On its official blog Mistral writes: "We will release the weights by the end of the month"; The Next Web gives 27 October as the expected date. The licence, however, has not been specified, leaving open whether the Apache 2.0 terms of its predecessor, Mistral Large 3, will carry over. On pricing, the official list rate of $1.36 per million input tokens and $4.18 per million output tokens currently comes with a 50% discount whose end date has not been disclosed.

The company's narrative leans heavily on cybersecurity and coding. According to figures released by Mistral, which the company itself describes as preliminary and subject to change because reinforcement learning is still under way, Large 4 scores 82% on a vulnerability reproduction and patching test, while Claude Opus 5.5 and GPT-6 Astra would score close to zero because they refuse the task: a company figure, not verified by third parties. In a blind coding evaluation run by Surge AI, Mistral reports a score of 3.74 out of 5 for Large 4, second of five models, behind Claude Opus 5 (4.22). Mistral AI's co-founder and chief scientist said that "cyber defense capabilities will enable enterprises and governments to defend themselves", previewing a version with reduced moderation for specialists and authorities. Still, without external, independent validation of these figures, the usual methodological caution applies.

According to Mistral, the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers; The Next Web instead mentions around 4,000 GPUs, about two months of training and roughly 10 megawatts of power. This infrastructure underpins the model's localisation, trained on more than 160 languages, and a future option for deployment managed entirely within the borders of the European Union. Presenting the launch, Mistral AI's CEO staked out a strategic place for the continent: "So the narrative that Europe cannot compete is something that is not true".

Technological sovereignty is also measured by how open the promised weights turn out to be. Much will depend on two things still unresolved: the actual release of the weights and the licence that comes with them, which has not yet been announced. Only then will we be able to tell how open Europe's first trillion really is. — Olya