Mistral has launched a public preview of Mistral Large 4, a natively multimodal model containing 1 trillion parameters, with 49 billion active parameters. Developers can try its API through Mistral Studio today, and Mistral plans to release the weights by the end of October. API pricing has not been disclosed.
Mistral says organizations will be able to deploy the model on their own premises or in private clouds, including for security operations they want to run under their own policies. Before releasing the weights, it is testing the model with cybersecurity specialists, vetted partners and state authorities, giving those participants fewer moderation restrictions and broader cybersecurity capabilities. The company plans availability in several regions, including a European deployment it will operate independently under European law.
Cybersecurity results include an 82% score on a test requiring the model to reproduce a vulnerability in open-source software and repair it, which Mistral says was the highest result among models tested. The company also reports that Mistral Large 4 completed 93% of Cybench's 40 security competition exercises. Its internal testing found uses for the model in malware analysis, vulnerability prioritization and the creation of detection rules.
For coding, Mistral reports a combined Coding Agent Index result of 49.8%, placing the model above DeepSeek V4 Pro 0813 and Qwen3.8 Max. Artificial Analysis evaluated the underlying benchmark results privately, ahead of a public launch of the testing harness that is still pending. In a separate blind assessment that Mistral says it ran with Surge AI, professional annotators placed ML4 Preview second among five models, scoring it 3.74 out of five compared with Claude Opus 5's 4.22.
Mistral reports a 59.9% result on AutomationBench, which tests 657 business workflows involving applications including Salesforce and Gmail. Its visual capabilities include inspecting technical drawings, finding evidence within PDFs and examining large satellite images, according to the company. On the Dense 200 visual grounding benchmark, Mistral reports a 42% result against GPT-6-Astra's 41%.
Training used 3,800 NVIDIA Grace Blackwell GPUs at Mistral's European datacenters, which also host the public preview. The training data included material across more than 160 languages, and the model uses the training, customization and reinforcement learning environment offered through Mistral Forge. Mistral also plans to use Mistral Large 4 as the basis for further specialized and optimized models.
