Mistral releases open weights for its 1 trillion-parameter Large 4 model

Mistral has published the weights of Mistral Large 4, a 1 trillion-parameter multimodal model that activates 52 billion parameters per task, on Hugging Face. The company calls it the strongest open model for agentic coding and cyber defence, with Terminal-Bench 4.0 rising from 28% to 41% and CyberGym-E2E at 79%.

• Model activates 52B of its 1T parameters and handles text and images
• Terminal-Bench 4.0: 41% versus 28% at preview, just ahead of GLM-5.3 at 40%
• CyberGym-E2E: 79%, compared with 76% for DeepSeek V4.1 Flash
• Trained on 3,800 Nvidia Grace Blackwell GPUs in Europe, RL run in France at ~10 MW