Mistral Large 4 preview: a 1T-parameter open-weight model pitched on refusal-free cyber work, weights due end of month

Why it matters

Key points

On October 6 Mistral released a public preview of Mistral Large 4 (nicknamed "le Chonk"): a natively multimodal model with 1 trillion total parameters and 52 billion active, its largest and most capable to date. The preview API is live on Mistral Studio, with weights planned for the end of October.

Mistral says the model excels at coding, agentic workflows and multimodal understanding, competes with the strongest open models worldwide, and clearly outperforms every open-weight model built in the US or Europe. It claims state-of-the-art results among open models on enterprise workloads such as cybersecurity, finance and law, and says its visual grounding even beats frontier closed models.

Cybersecurity is the main pitch. Mistral says ML4 ranks in the global top five on the Artificial Analysis Cyber Index and leads open-weight models built outside China by a wide margin. On one test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scores 82%, the highest of any model, and it solves 93% of the 40 Cybench challenges. Mistral notes that Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task.

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, which also serve the preview. Training data spans more than 160 languages, including every official EU language. Before releasing weights, Mistral is red-teaming the model with cybersecurity firms, vetted partners and state authorities, who get the same model with lighter moderation and expanded cyber capabilities.

Verifiable

Official blog: https://mistral.ai/news/mistral-large-4/ , published October 6, 2026 at 12:00 UTC per the page metadata.

All scores are self-reported or cited by Mistral; it says architecture details, more benchmarks and post-training methods will follow before the weights ship. The page does not list every model behind "the strongest open models worldwide." The near-zero scores for closed models reflect refusal policies, not capability gaps. The end-of-month weight release is a plan, not yet a fact.

Our take

Large 4 gives Europe its first trillion-parameter model in the top tier of open weights, but the notable part is the pitch Mistral chose: cyber capability without refusals, plus self-hosting. That openly treats closed labs' safety filters as a product flaw, and for defenders it is a real benefit, since proving a vulnerability exists is often exactly the step that gets refused.

The risk sits in the same place: those capabilities serve attackers just as well, and open weights cannot be recalled. Red-teaming with loosened moderation before releasing the weights is a reasonable sequence, but whether the weights ship on time, and with what safeguards and license terms, is the real test of Mistral's sovereign-AI story. Until then it is a preview you can only try through Mistral's own API.