On October 29, 2026, Mistral AI released the open weights of Mistral Large 4 (Le Chonk) through Hugging Face. It is the company's largest and most capable model to date: about 1 trillion parameters in total, 49B active per token, a 1M-token context window, and text plus image input.
Unlike the public preview on October 6, the point of this release is that the weights themselves can be downloaded and self-hosted. Official FP8 and NVFP4 variants are available, together with a built-in DSpark speculative decoding draft head. Hugging Face's launch post says transformers, vLLM and llama.cpp all supported the model on day one, and vLLM reported that FP8 and NVFP4 stayed within 2.1 points of each other across several evaluations.
Security is the headline capability
Mistral Large 4 scored 82% on a test that reproduces a real vulnerability and then patches it, and solved 93% of Cybench, a set of 40 security-competition challenges. Hugging Face's post notes that closed models often refuse to reproduce vulnerabilities because of safety filters, while open weights let defense teams control deployment and usage policy themselves. That does not mean the model accepts everything: its average refusal rate on jailbreak and harmful-request tests was higher than that of other open-weight models.
Downloadable does not mean runnable for everyone
The bar for running it locally remains high. Unsloth's 1-bit GGUF quantization shrinks the model to about 233GB and is said to retain 91% top-1 accuracy, yet it still needs more than 256GB of memory and targets Mac Studio-class machines. More typical deployments still call for eight Blackwell-class GPUs.
The license has limits too. The weights use the Mistral Ecosystem Open-Weight License, a modified MIT license: use, modification, fine-tuning and commercialization are free, but offering the model as a service requires a separate commercial license from Mistral. The Mistral API charges $0.70 per million input tokens and $2.10 per million output tokens.
For security, finance and legal teams, the value of this release is not the parameter count itself. It is that a frontier model with standout security results can now live inside their own infrastructure. Before installing it, though, they need to work out the hardware, licensing and operations costs.
