Mistral has released the weights for Mistral Large 4 ("Le Chonk"), a trillion-parameter mixture-of-experts model with 52 billion active parameters, under a modified MIT license requiring a separate agreement for commercial cloud hosting. The model accepts text and image inputs and outputs text and code, running on eight Nvidia B300 GPUs via SGLang or four GB300 GPUs in a compressed variant. Following three weeks of post-preview reinforcement learning on progressively harder in-house tasks, Large 4 improved its Terminal-Bench 4.0 score from 28% to 41%, edging out GLM-5.3 (40%) and tying GLM-5.3 at 48% on Artificial Analysis' Coding Agent Index while using 47% fewer tokens. However, it trails Claude Opus 5.5 (62% on Coding Agent Index, 46% on FrontierCode 2.0 vs. Large 4's 19.7%) and the upcoming Kimi K3.1 Pro (67% on Terminal-Bench). On Artificial Analysis Intelligence Index v5, Large 4 scores 26 points, behind DeepSeek V4.1 Flash and GLM-5.3 (32 each) and MiMo-V2.6-Pro (30), with the preview gaining only one point from extra training. In cyber defense, Large 4 scores 51% on the Cyber Index, ranking third among open models behind DeepSeek V4.1 Flash (57%) and MiMo-V2.6-Pro (56%), though it leads in refusing harmful cyber requests at 96.6%. API pricing is $0.70/M input and $2.10/M output tokens; index task cost averages $2.06, cheaper than GLM-5.3 ($5.01) and Kimi K3.1 Pro ($3.68) but above DeepSeek V4.1 Flash ($0.60). Generation speed reaches 161 tokens/second, and total index output tokens dropped from 150M to 110M, nearing the 100M median for comparable models.