Mistral releases the weights of its trillion-parameter Large 4 model and warns it could help with real attacks

Mistral has posted the weights of Mistral Large 4, nicknamed "Le Chonk," on Hugging Face, three weeks after a public preview. The multimodal model has about one trillion parameters, 52 billion of them active per step, and is released under a modified MIT license. A quantized version runs on four Nvidia B200 or B300 GPUs. Mistral says three more weeks of reinforcement learning improved the model. On Terminal-Bench 4.0, its score rose from 28% at preview to 41%, just ahead of GLM-5.3 at 40%. The company says it matches GLM-5.3 on Artificial Analysis' Coding Agent Index while using 47% fewer tokens per task, and it scored 79% on the CyberGym-E2E bug-patching test. On the broader Artificial Analysis Intelligence Index, The Decoder reports a score of 26, behind DeepSeek V4.1 Flash and GLM-5.3 and far behind the leading Anthropic and Google systems. Its strengths include a speed of 161 tokens per second and a 96.6 percent refusal rate on harmful cyber requests. Mistral wrote that although the preview refused black-box attacks on live targets, "we see a potential for uplift to conduct real attacks." It says it will keep updating the guardrails. CSO Online notes that those updates reach only Mistral's API and hosted services, and Mistral has not said how they would apply to weights already downloaded.