Mistral previews Large 4, a 1-trillion-parameter model whose weights it will publish this month
Victoria Shinder1 min readMistral AI announced a public preview of Mistral Large 4 on 6 October. The preview API is available now on Mistral Studio, and the company says it will publish the weights by the end of the month.
The model. Large 4 is a natively multimodal model with 1 trillion parameters, of which 49 billion are active. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe, that the preview runs on the same infrastructure, and that more than 160 languages were in the training data, including every official EU language.
Why the weights are delayed. Until release, Mistral says it is red-teaming the model with cybersecurity firms, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities.
The numbers Mistral reports. On one test of the Artificial Analysis Cyber Index, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Large 4 scores 82%, which Mistral says is the highest of any model; it solves 93% of the 40 Cybench challenges. On coding it reports 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. In a blind human evaluation of coding quality run with Surge AI it ranked second of five models, behind Claude Opus 5. Mistral's own explanation for the cyber gap is that several closed models refuse the vulnerability-reproduction task and so score near zero on it.
All of these figures are the vendor's own. Mistral says it will share architecture details, more benchmarks and its post-training method as the weights approach release; that is the point at which independent tests become possible.