Mistral AI

AI · Open-Source AI · Cybersecurity

Launch

Mistral releases open weights for trillion-parameter Large 4

Enterprises and governments can now self-host a frontier coding and cyber-defense model without depending on a foreign AI vendor.

• Mistral AI released open weights for Mistral Large 4, nicknamed "Le Chonk," a 1-trillion-parameter multimodal model, three weeks after its public preview debut.
• Mistral, a Paris-based AI lab competing with OpenAI, Anthropic and Chinese open-weight labs, trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers.
• The mixture-of-experts model activates just 49-52 billion of its 1 trillion parameters per token, reads text and images, and supports a 1-million-token context window across 160+ languages.
• Three extra weeks of reinforcement learning lifted Large 4's Terminal-Bench 4.0 coding score from 28% to 41%, edging past Chinese rival GLM-5.3 at 40% while using 47% fewer tokens per task.
• On cybersecurity, Large 4 scored 79% on CyberGym-E2E for finding and patching real memory-safety bugs, beating DeepSeek-V4.1 Flash's 76% and matching MiMo-V2.6-Pro.
• Mistral itself warned the model could offer "uplift to conduct real attacks," and later safeguard updates will only reach its hosted API, not copies already downloaded.
• The UK AI Security Institute previously found open-weight models lag closed frontier systems by four to seven months on cyber tasks, warning that releasing highly cyber-capable weights creates irreversible misuse risk.
• By open-sourcing a frontier-grade coding and cyber-defense model, Mistral sharpens the case that sovereign, self-hosted AI can rival closed US and Chinese labs on capability while undercutting them on cost and control.