What happened

Mistral published an October 29, 2026 post saying the weights of Mistral Large 4 (Le Chonk) are available after a public preview earlier in October. The Hugging Face card describes a roughly 1 trillion parameter Mixture-of-Experts model with about 49 billion active parameters per token (52 billion including embeddings and output layers), 128 routed experts with four active per token, a 1 million token context, and text-plus-image input. Checkpoints include FP8 and NVFP4. Mistral says three weeks of further reinforcement learning after the preview pushed open-weight coding results: 19.7 percent on Cognition's private FrontierCode 2.0, 48 percent on the Artificial Analysis Coding Agent Index while using 47 percent fewer tokens per task than GLM-5.3 on that index, and 41 percent on Terminal-Bench 4.0, up 13 percentage points from the preview with 30 percent fewer output tokens. The post adds Low, Medium, and High reasoning modes. On Artificial Analysis Cyber Index, Mistral reports 51 percent, and an average 96.6 percent refusal on its malicious cyber request set. The license is Mistral's Ecosystem Open-Weight License; Model-as-a-Service use needs a separate commercial license.

Why it matters

Open-weight coding and cyber claims now come with a downloadable 1T-class European MoE, not only a hosted preview. Teams that need to self-host, inspect, or fine-tune a frontier-sized open model have a dated primary page with concrete scores. Those scores are still Mistral's and Artificial Analysis's published figures, not an independent rerun of every table.

Mistral says the model was trained from scratch on European infrastructure, including 3,800 NVIDIA Grace Blackwell GPUs on the Hugging Face card.

A specialized cybersecurity variant for vetted enterprise design partners is still preview-access, separate from the public open weights.

Mistral reports 26 on Artificial Analysis Intelligence Index v5, calling it the highest among western open-weight models on that aggregate while noting Chinese open-weight leaders score higher overall.