Mistral previews 1T-parameter open-weight Large 4
Mistral has opened a public API preview of Mistral Large 4, a 1-trillion-parameter multimodal model with weights due by month's end, and is pitching its low-refusal cybersecurity performance as the main advantage over closed models.
Mistral has launched a public preview of Mistral Large 4 (ML4). The model is available now through the Mistral Studio API, and the company says it will release the weights by the end of the month.
ML4 is a natively multimodal model with 1 trillion total parameters and 49 billion active parameters. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and the preview runs on the same hardware. The company says more than 160 languages made up a significant share of the training data.
Benchmarks. All figures are Mistral's own claims. On agentic coding, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, using Artificial Analysis numbers. Its Coding Agent Index score of 49.8% puts it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind coding-quality evaluation run with Surge AI, it scored 3.74 out of 5, second to Claude Opus 5 (4.22) and ahead of Kimi K3 and GLM-5.3.
Cyber positioning. Mistral says ML4 ranks in the top five on the Artificial Analysis Cyber Index and solves 93% of Cybench challenges. It scores 82% on a test where models must reproduce a real vulnerability and then patch it. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task, and it presents open weights and self-deployment as a way around provider-level refusals.
Before the weights ship, vetted partners, cybersecurity firms and state authorities are red-teaming a version of the model with reduced moderation. Mistral says it will publish the architecture, more benchmarks and its post-training methods later. The company concedes ML4 is still a preview that is "continuing to improve."