Mistral today released the weights for Mistral Large 4 (Le Chonk), giving developers an improved version of its 1-trillion-parameter multimodal model that they can customize and host themselves. The release introduces three reasoning modes and coding improvements that Mistral says reduce the number of tokens needed for tasks, lowering inference costs. The model had entered public preview earlier this month, before another three weeks of reinforcement learning.
Developers can select Low for simpler work, Medium for routine tasks or High for demanding problems requiring more precise answers. High is the default setting, while Low takes the place of the previous None reasoning option. Mistral says it trained the model from scratch using infrastructure in Europe and offers deployment under organizations' own control.
Read more: Mistral Launches Mistral Large 4 Preview With 1 Trillion Parameters
Mistral reports a 48% result on the Artificial Analysis Coding Agent Index, roughly level with GLM-5.3, while consuming an average of 47% fewer tokens for each task. On Terminal-Bench 4.0, the score rose from the preview's 28% to 41%, with output token usage falling 30%, according to the company. Mistral says that result puts the model ahead of its open-weight competitors on the terminal coding benchmark.
The company also reports 19.7% on Cognition's private FrontierCode 2.0 benchmark, which tests code quality across a complete task and whether a maintainer would accept the resulting repository change. In blind comparisons run by Surge Code, evaluators favored Mistral Large 4 over GLM-5.3 and Kimi K3, Mistral says. However, Claude Opus 5.5 placed ahead of it in those evaluations.
For defensive cybersecurity, Mistral reports a 51% score on the Artificial Analysis Cyber Index. That assessment tests whether models can find security flaws, reproduce and verify them, then make repairs while preserving existing functionality. The company says it has also worked with cybersecurity vendors since the preview to assess vulnerability reproduction and patching in practical environments.
Mistral says the model refused malicious cybersecurity requests at an average rate of 96.6%, the highest among the open models it measured. Its refusal testing used cybersecurity prompts drawn from AgentHarm, JailbreakBench and StrongREJECT. The company attributes that behavior to guardrails intended to distinguish legitimate defensive work from malicious offensive tasks.
