Mistral AI announced on October 6, 2026 the public preview of Mistral Large 4, or ML4, nicknamed "the Chonk." The model has 1 trillion parameters, 49 billion of them active, is natively multimodal, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company's European data centers. The preview API is available on Mistral Studio; the weights will be released before the end of the month, following a real-world testing phase with cybersecurity leads, verified partners, and state authorities, who get access to a version with reduced moderation and extended cyber capabilities.
Mistral presents ML4 as close to the world's best open models and clearly ahead of any American or European open model, with state-of-the-art results among open models in cybersecurity, finance, and law, and results that surpass some closed models in visual grounding.
In cybersecurity, the model ranks in the top five of the Artificial Analysis Cyber Index and leads open models outside China. On the vulnerability reproduction-then-fix test, it scores 82%, the highest score cited, and 93% on Cybench. The post explains that Claude Opus 5.5 and GPT-6 Astra score close to zero on this test because they refuse the task, whereas defending software often starts with proving that a flaw is real. It concludes that this shows the value of open models, deployable privately or on-premises, under the organization's own policy.
In agentic code, ML4 reaches 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined index of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind human evaluation by Surge AI ranks it second out of five (3.74), behind Claude Opus 5 (4.22). In agentic workflows, it scores 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. On knowledge work, vals.ai ranks it ahead of GPT-6 Astra in finance and law, and it outperforms other open models on Harvey's legal benchmark. The security section cites 93.3% of attacks blocked on Lakera's B3 benchmark and a higher refusal rate for malicious cyber requests than the open models it is compared against.
On methodology, Mistral describes a composable reinforcement-learning library in which sandboxes, web search, APIs, and verifiers (reward models, tests, judges) combine. With 3,000 GPUs, a training run produces about 33 billion tokens per day. ML4 is the first step of the roadmap funded by the série D de 3 Md€ round and will serve as the base for specialized models. Pricing is $1.36 per million input tokens and $4.18 per million output tokens.