Mistral Large 4 open weights arrive on Hugging Face with multiple deployment options

Mistral Large 4, which the post calls Le Chonk, takes text and images as input and generates text. It combines a large pool of experts with sparse attention and a Pixtral vision encoder. The model is distributed under the Mistral Ecosystem Open-Weight License, which the post describes as an MIT derivative that sets limits on model-as-a-service offerings.

The post reports 82% on the Artificial Analysis Cyber Index vulnerability reproduction and patching test and 61.7% on DeepSWE v1.1. It says the model is the fifth best open weights model on the Artificial Analysis Intelligence index as of October 29, 2026. Cybench results of 93% and 93.3% attack resistance on Lakera's B3 benchmark are also listed.

Developers can run the weights with transformers, with vLLM using the NVFP4 checkpoint on NVIDIA Blackwell GPUs, or with llama.cpp and GGUF quants. The model is also available on Inference Endpoints with a recipe for 8 x NVIDIA H200 GPUs, and TRL supports fine-tuning. Dell Enterprise Hub offers a deployment recipe for NVIDIA B300 nodes.

Teams building coding or security agents can self-host a model with image input and a 1M-token context window, rather than relying only on a hosted API that may block security work. The post notes that reasoning cannot be disabled, so setting reasoning_effort to none still uses high. Hardware requirements are significant, since the full checkpoints need multi-GPU setups, while the 2-bit GGUF is described as running on a single Apple Silicon Ultra with 512 GB of unified RAM.