Mistral Large 4 went into public preview on October 6. On October 29 Mistral shipped an improved version and put the weights on Hugging Face in FP8 and NVFP4, with GGUF quants from Unsloth and the llama.cpp team. Mistral's docs list 1.05T total parameters, 52B active and a 1.6B vision encoder (Hugging Face's post says 49B active; the Transformers release notes explain the gap: 52B counts embeddings and output layers). It takes text and images, and the context is 1M tokens.
Mistral's numbers: 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task, 41% on Terminal-Bench 4.0 (up 13 points from the preview) and 19.7% on Cognition's private FrontierCode 2.0. There are now three reasoning modes: Low, Medium and High, which is the default.
On the API it costs $0.70 per million input tokens, $0.07 cached and $2.10 output. It's also on OpenRouter, Microsoft Foundry and NVIDIA endpoints, and Transformers 5.22.0, vLLM and llama.cpp support it from day one. The license lets you use, fine-tune and sell products on it. The catch is section 2: if you run a Model-as-a-Service business, meaning you give third parties inference access with real control over inputs, you need a commercial license from Mistral first. Internal use is exempt.