Mistral AI announced the open-weight release of Mistral Large 4, nicknamed Le Chonk, on October 29, 2026; Hugging Face published its official introduction the same day. Unlike the October 6 API public preview, this event makes the improved model downloadable for deployment on an organization's own infrastructure. The official card links FP8 and separate NVFP4 checkpoints. It is a large mixture-of-experts model accepting text and images and producing text. An open-weight release is not blanket general availability across every service, or an unlimited free hosted product.
Read the deployment freedom and license together
The Mistral Ecosystem Open-Weight License is a modified MIT-style license permitting use, modification and further training. It also imposes a separate commercial-license condition on defined Model-as-a-Service businesses using the model commercially; its definition and internal-use exception matter. Calling it unrestricted MIT or Apache 2.0 would be inaccurate. AIZIGOO's interpretation is that teams wanting less API dependency and a model adapted to internal code or documents gain another option. They also assume responsibility for data access, retention, tool permissions and model updates.
Downloadable does not mean lightweight or cost-free
Official deployment examples use substantial memory and multiple GPUs, while the NVFP4 route targets Blackwell hardware. Quantization and sparse activation do not establish ordinary-laptop performance at the same speed or context length. Hardware, power, storage and operations still cost money, and hosted API inference is separately paid. Coding and security results described by Mistral and Hugging Face vary with the checkpoint, evaluation setup and date; this article did not independently reproduce the rankings. Our adoption suggestion is to test patch correctness, review time, end-to-end latency and cost on scoped repository samples with restricted permissions. A security benchmark or refusal rate is not a guarantee against misuse or unsafe deployment.