Mistral Large 4 is a 1T-parameter mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → (49B active), multimodalA model that works with more than text — reading images, audio, or video, and sometimes generating them too.Full definition →, with preview API access now and open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → due by end of October. Engineers choosing open models for coding and agents get a new candidate to evaluate.
Key takeaways · AI-distilled
Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, the preview runs on that same infrastructure, and a European deployment will operate under European law.
Until the weights ship, Mistral is red-teaming ML4 with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.
Mistral reports ML4 scores 82% on a test that has models reproduce and then patch a real vulnerability, and says several closed models score near zero there because they refuse the task.
For coding, Mistral cites 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, and a blind Surge AI human rating that placed the preview second of five (3.74), behind only Claude Opus 5 (4.22).
Mistral says architecture details, more benchmarks and its post-training method will be shared as the weights approach, so for now the results are vendor-reported, with some coding scores drawn from Artificial Analysis.
Terms in this piece · Glossary
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
