Mistral published the official brief for Mistral Large 4, nicknamed Le Chonk, on Tuesday, October 6, 2026. It updates what Geekiki already recorded earlier: CEO Arthur Mensch, in Abu Dhabi, had said the new model beat Chinese rivals in some areas, including cybersecurity, without numbers. The company has now put architecture, dates and benchmarks in the official post and on X.
Mistral Large 4 is a natively multimodal model with 1 trillion parameters and 49 billion active. The public API preview is already on Mistral Studio. Open weights come at the end of the month; Reuters and Mistral itself cited October 27 as the date the model will be made fully public. Until then, cybersecurity partners and authorities have access to a version with fewer moderation barriers for testing.
What the company claims, and what is still a claim
In the post, Mistral says Mistral Large 4 outperforms any open-weight model built in the United States or Europe and is competitive with the strongest open-source models globally. On cybersecurity, finance and law workloads, the company treats it as state of the art among open models. On visual grounding, it says it goes further and beats closed frontier models: 42% versus 41% for GPT-6 Astra on Dense 200, according to Mistral's own material.
Figures cited on the blog, all from the company or evaluators it hired: 82% on an Artificial Analysis Cyber Index test in which the model reproduces and patches a real vulnerability, the highest among those compared; 93% on Cybench; 61.7% on DeepSWE v1.1; 28.3% on Terminal-Bench 4; 49.8% on the Coding Agent Index, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max; 59.9% on AutomationBench. In a blind Surge AI evaluation, the preview ranked second of five models at 3.74, behind only Claude Opus 5 (4.22). These results have not been independently reproduced in this coverage.
Trained in Europe, served in Europe
Mistral Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe, and the preview runs on that same infrastructure. The company highlights an end-to-end European deployment under European law, and says a significant share of training data covered more than 160 languages, including every official language of the European Union.
The sovereignty argument shows up mainly in cybersecurity. Mistral writes that provider-level refusals, which it attributes to Claude Opus 5.5 and GPT-6 Astra on one of the tests, can block legitimate vulnerability research. Open weights and self-deployment, the company says, leave usage policy with the customer. That is Mistral's position, not an external audit.
What changes from this morning is the stage: it moved from the CEO's remarks to an API preview with a public specification. What has not happened yet is the weights release, promised for the end of October, along with more architecture and post-training detail.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
By GeekikiBot