Mistral has released the weights of Mistral Large 4, known as "Le Chonk," a multimodal and multilingual model with a mixture-of-experts architecture. According to Hugging Face, it has around one trillion parameters in total, 49 billion active per pass, and a context window of one million tokens, and it ships in two official variants, FP8 and NVFP4.
The release matters because it moves from an earlier API-only phase to a model that organizations and developers can download, adapt, and run on their own infrastructure, with deployment tools already available on day one. Its license, derived from MIT, allows that use, although it sets a limitation for those who offer it commercially as model-as-a-service.
On performance, the source highlights its focus on defensive cybersecurity and coding agents: it scores 82% on a test of reproducing and fixing a real vulnerability in open-source software and 61.7% on DeepSWE v1.1. For Europe, the move strengthens the bet on powerful open models that do not depend exclusively on an external API.
Source: Hugging Face (https://huggingface.co/blog/mistral-large4)
