Mistral makes its trillion-parameter Large 4 weights public with focus on coding and defense

French artificial intelligence firm Mistral has officially published the weights for Mistral Large 4, a trillion-parameter multimodal model nicknamed Le Chonk, following an initial preview release on Oct. 6. Available under the Mistral Ecosystem Open-Weight License, the release includes official FP8 and NVFP4 variants alongside community GGUF formats. Mistral reported that the model was trained from scratch across two months using 4,000 Nvidia Grace Blackwell GPUs in its European data centers. [2][4]

Mistral Large 4 operates with about 1T total parameters and 49B active parameters across a 1M-token context window, combining sparse attention with a Mixture-of-Experts architecture and a Pixtral vision encoder. The model features low, medium, and high reasoning modes. On the Terminal-Bench 4.0 coding benchmark, its score reached 41%, compared to 28% during preview testing. Mistral claims the system matches rival open models such as Opencode's GLM-5.3 while consuming 47% fewer tokens per task on average. [2][4]

Mistral is heavily pitching the model for defensive cybersecurity, arguing that open-weight architectures give organizations data sovereignty and prevent outages during incident response. Mistral Large 4 scored 82% on the Artificial Analysis Cyber Index test for reproducing and patching software flaws, placing it among the top five open models. Mistral claims this release leapfrogs leading Chinese models like DeepSeek in key areas, though its license does introduce usage limits for third parties offering model-as-a-service platforms. [2][3][4]