Mistral has released the weights for Mistral Large 4, its one-trillion-parameter multimodal model nicknamed Le Chonk. Organizations can now customize and run the model on their own infrastructure rather than relying only on a hosted public preview.
The model processes text and images and is built as a mixture of experts, activating only part of its total parameter set for each token. Mistral says three additional weeks of reinforcement learning improved coding, defensive cybersecurity and reasoning efficiency. On Cognition’s private FrontierCode 2.0 benchmark, the company reports a 19.7 percent score, leading the open-weight systems it compared.
Those figures remain vendor claims tied to specific tests. Independent results cited by other coverage show that Large 4 is fast and competitive on coding while still trailing leading closed systems and some Chinese open models on broader intelligence measures. Open weights also do not mean lightweight deployment: even compressed variants require multiple high-end GPUs.
The release matters most to governments and companies that need to inspect, customize or self-host a capable model. It also preserves a dual-use trade-off. The same access that helps defenders analyze malware and patch vulnerabilities can give attackers a model that cannot be disabled remotely.