Mistral releases weights for Mistral Large 4 model

Mistral released the weights for Mistral Large 4, nicknamed 'Le Chonk', on Hugging Face. The company made the model available three weeks after its public preview. Mistral Large 4 is a multimodal, multilingual, Mixture-of-Experts model that takes text and images as input and generates text. The model is distributed under the Mistral Ecosystem Open-Weight License.

The model features about one trillion total parameters with 49 billion active parameters. It includes a context window of one million tokens. According to SGLang documentation, the standard version of the model runs on eight Nvidia B300 GPUs. A more heavily compressed variant runs on four GB300 GPUs. Mistral researchers Pierre Stock and Corentin Petit walked through the training process on a press call. They noted that the model's score on Terminal-Bench 4.0 rose from 28 to 41 percent after three more weeks of training.

In internal testing, the model proved useful for malware analysis, vulnerability prioritization, and writing detection rules. Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test. The model has an average refusal rate higher than any other open weight model on specific cyber prompts. Mistral claims its model outranks all other open models on agentic coding tasks on Artificial Analysis' Coding Agent Index, where it hits 48 percent. The model uses 47 percent fewer tokens than GLM-5.3 to achieve similar coding performance.

Mistral CEO Arthur Mensch stated that AI should be a tool for empowerment, not dominance. Pierre Stock stated that organizations can access frontier intelligence to defend themselves against cyberattacks. Moonshot plans to release Kimi K3.1 Pro weights by November 4, 2026.