Mistral Large 4 opens as a public preview: 1 trillion parameters, 49 billion active, and 82% on a cyber vulnerability test
Synopsis
Mistral released a public preview of Mistral Large 4, a natively multimodal model with 1 trillion total parameters and 49 billion active parameters, with weights promised by the end of the month. Open-weight models have generally trailed the strongest closed frontier models on enterprise tasks such as cybersecurity, finance and law; Mistral says ML4 is state of the art among open models in those areas and edges GPT-6-Astra on the Dense 200 visual-grounding test, 42% versus 41%. On one Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability and then patch it, ML4 scores 82%, the highest of any model.
Interpretation
Mistral put Mistral Large 4 into public preview and committed to releasing its weights by the end of the month, bringing a trillion-parameter-class multimodal model into the self-deployable open-weight line. Open-weight models have generally been seen as trailing the strongest closed models on critical enterprise workloads such as cybersecurity, finance and law; Mistral says ML4 reaches state of the art among open models in those areas. On an Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82%, the highest of any model; on the same test Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.
The model unifies instruction, reasoning and agentic capability in one mixture-of-experts architecture that accepts both text and images. Mixture of experts means only part of the parameters activate on each forward pass, so the model has 1 trillion total parameters but 49 billion active. Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and the preview API is served on that same infrastructure.
On agentic coding and general workflows, ML4 is compared head to head with contemporary open models. Mistral reports a combined Coding Agent Index score of 49.8%, which it says places ML4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. It scores 59.9% on AutomationBench, 657 business workflows, and reaches 1,393 Elo on AA-Briefcase, a long-horizon knowledge-work evaluation, both reported above DeepSeek V4 Pro.
Perspective
This is a public preview; the weights are planned for release by the end of the month, with architecture details, additional benchmarks and post-training methodology to follow then. During the preview, Mistral is red-teaming the same model with reduced moderation and expanded cyber capabilities alongside cybersecurity leaders, vetted partners and state authorities in real-world settings.
Mistral says the reinforcement learning run behind the preview is still in flight and the model shows no signs of saturation, with large improvements expected in the weeks and months ahead; most benchmarks are vendor-reported, some comparisons were run by Mistral itself, and the weights are not yet out.