Welcome Le Chonk: Mistral Large 4 on Hugging Face

Hugging Face has published a launch post, dated October 29, 2026, for Mistral Large 4, an open-weight multimodal model nicknamed Le Chonk. The Mixture-of-Experts system accepts text and images, generates text, and has roughly 1 trillion total parameters, about 49 billion of them active, with a 1-million-token context window. It ships in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License, an MIT derivative that restricts model-as-a-service offerings.

Cybersecurity is the headline pitch. The post says the model scored 82% on an Artificial Analysis test that requires reproducing and then patching real open-source vulnerabilities, the highest reported result on that task, while Mistral claims several closed frontier models score near zero because they refuse. It also reportedly solves 93% of 40 Cybench challenges. Mistral frames the model as a tool for defenders, and the post says its average refusal rate on harmful cyber prompts exceeds that of any other open-weight model.

Elsewhere, Hugging Face ranks it fifth among open-weight models on the Artificial Analysis Intelligence index as of publication, with a 49.8% Coding Agent Index score that beats DeepSeek V4 Pro 0813 and Qwen3.8 Max.

Support is available from day one in Transformers, vLLM, llama.cpp and TRL, plus Inference Endpoints and Dell Enterprise Hub. Running it remains demanding: the 2-bit quantizations fit a single Apple Silicon Ultra machine with 512 GB of memory, where Hugging Face measured about 15 tokens per second, while other versions need distributed inference.