Mistral Large 4 preview live; open weights due late October

Mistral AI's move puts its largest model, Mistral Large 4, into developers' hands through Mistral Studio before the open weights arrive. This is a deliberate sequence: the company gets feedback and real-world testing while retaining control over the exact release moment. The model's nickname, "Le Chonk," and its chubby-cat mascot signal that Mistral is comfortable with a playful public identity even for a 1.05 trillion-parameter system.

What distinguishes this release is Mistral's emphasis on cybersecurity. The company claims the model tops a vulnerability reproduction and patching test at 82%, and it argues that closed models like Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse such tasks. Mistral is positioning open weights as the tool that lets defenders independently verify exploits — a stance that may appeal to security researchers who chafe at refusal-based guardrails. The company is also running red-teaming with cybersecurity firms, vetted partners, and state authorities before the release, an unusual pre-release step that may shape the final weights.

Upcoming milestones carry real stakes for whether developers treat Mistral Large 4 as production-ready. The open-weight release at the end of October will determine whether independent teams can audit and self-host the model, and the results of the pre-release red-teaming may indicate how much moderation the company keeps in the public version. For now, the preview offers a usable API but leaves the question of long-term openness unanswered.