French AI lab Mistral on October 6 released Mistral Large 4, its largest and most powerful model to date, internally codenamed "Le Chonk" (meaning "big fat cat"), mounting a direct challenge to US closed frontier models and Chinese open-weight models.
Point one: a trillion-parameter mixture-of-experts architecture. The model has 1 trillion total parameters with only 49B active per inference, is natively multimodal, supports more than 160 languages and has a context window of up to 1 million tokens. A public API preview is now live.
Point two: efficient training in Europe. The model was trained from scratch in Mistral's own European data centers on about 3,800 Nvidia Grace Blackwell GPUs, reportedly using only one-half to one-third of the training compute of Chinese rivals.
Point three: standout cybersecurity capability. It scored 82% on the CyberGym-E2E benchmark and 93% on Cybench, while closed models such as Claude and GPT scored close to zero on that test because they refused to answer. CEO Artur Mensch said the claim that "Europe cannot compete" does not hold. The full open weights are expected to be released at the end of October.
Source: read the original. This text is a summary for reference only.
