Mistral previews a trillion-parameter model ahead of its planned Oct. 27 weights release

French AI company Mistral has opened a preview of Mistral Large 4, its largest model yet, and calls it “Le Chonk.” The model has roughly one trillion total parameters—the internal settings learned in training—but uses a mixture-of-experts design that activates about 49 billion for each task, helping make a very large system more practical to run. Mistral says it beats every open-weight model developed in the US or Europe and is especially strong in code, cybersecurity, and image-based tasks; those are company claims, not a final independent verdict. The meaningful next step is October 27, when Mistral plans to publish the weights, giving organizations the option to run, adapt, and govern the model on their own infrastructure rather than depend solely on Mistral’s API.