Mistral AI has released the open weights for Mistral Large 4 ("Le Chonk"), a 1-trillion-parameter natively multimodal Mixture-of-Experts model trained from scratch on European infrastructure. Following three weeks of additional reinforcement learning post-preview, the model leads open-weight benchmarks in agentic coding and defensive cybersecurity. On the Artificial Analysis Coding Agent Index, it scores 48%—tied with GLM-5.3—while using 47% fewer tokens per task; on Terminal-Bench 4.0, it reaches 41% (+13 pp vs. preview) with 30% fewer output tokens; and on Cognition's private FrontierCode 2.0, it achieves 19.7%. In Surge Code blind evaluations, it ranks behind only Claude Opus 5.5 among tested open models. For cybersecurity, it scores 51% on the Artificial Analysis Cyber Index (vulnerability discovery, reproduction, and patching) and refuses 96.6% of malicious requests, the highest measured among open models, while resisting 97.1% of attacks on the B3 Agent Security Benchmark. Enterprise metrics include 73.3% on Finch (finance spreadsheets) and 84.5% on BrowseComp (web browsing). The model introduces three reasoning modes (Low, Medium, High/default) for cost-performance control. Available under a modified MIT license (Mistral Ecosystem Open-Weight) requiring a separate commercial license for Model-as-a-Service providers, API pricing is $0.70/M input, $0.07/M cached input, and $2.10/M output tokens. Deployment options include self-hosting via vLLM/SGLang, Hugging Face, Microsoft Foundry, NVIDIA NIM, and Mistral Studio/Forge for customization.