Mistral Large 4 preview: 1T-parameter open model, weights due this month

Mistral has opened a public preview of Mistral Large 4 (ML4) on its Mistral Studio API. It has 1 trillion parameters, of which 49 billion are active at a time, and reads text and images natively. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres. Mistral says the weights will be released by the end of October. Until then it is red-teaming the model with security firms, vetted partners and state authorities. Mistral's own numbers: 61.7% on DeepSWE v1.1 (coding), 59.9% on AutomationBench (657 business workflows across Gmail, Google Sheets, Slack and Salesforce), and 93% of the 40 Cybench security challenges. In a blind coding review run with Surge AI, it came second of five, behind only Claude Opus 5. Mistral also pitches it at security teams: on one test that asks a model to reproduce and patch a real vulnerability, ML4 scored 82%, while Claude Opus 5.5 and GPT-6 Astra scored near zero because they refused. These are vendor results and have not been independently checked yet. On Hacker News (about 1,560 points, 950+ comments), one camp is pleasantly surprised: better benchmarks than expected, strong vision and cyber numbers, and a real European open-weight contender. The other camp is cautious: Mistral's last big model ranked low on public leaderboards, some say this is still a generation or more behind the top closed models, and a few ask whether it leaned on distilling Chinese open models. The rough consensus: a serious release worth testing, but people want independent evals and the actual weights before calling it. Sorted take: if you need a model you can host yourself or keep inside the EU, put ML4 on your shortlist now. Test it on your own documents through the preview, then decide once the weights and independent results land.