Mistral AI has launched a public preview of Mistral Large 4 (ML4), its largest and most capable model to date. The model is natively multimodal, capable of reasoning across complex documents, charts, and natural images. While it features a total of 1 trillion parameters, it utilizes a Mixture-of-Experts (MoE) architecture with 52 billion active parameters. Mistral AI plans to release the model's weights by the end of October 2026, allowing organizations to deploy the model on private infrastructure for greater control over data and AI sovereignty.
ML4 demonstrates significant performance in specialized enterprise verticals, particularly in cybersecurity and coding. In cybersecurity, it ranks among the top five models globally on the Artificial Analysis Cyber Index, notably outperforming several leading closed models on specific tasks because it does not refuse to perform vulnerability research or patching. In coding, the model achieved a Coding Agent Index score of 49.8%, placing it ahead of competitors like DeepSeek V4 Pro and Qwen3.8 Max. Additionally, the model shows high proficiency in scientific workflows, such as generating full Hartree–Fock simulations, and excels in knowledge work, exceeding GPT-6-Astra on HarveyAI’s Legal Agent benchmark and FinWorkBench for financial analysis.
The model's performance in coding tasks is highlighted by its high score on the Coding Agent Index.
The article notes that ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index.
The model excels in knowledge work, specifically exceeding GPT-6-Astra on HarveyAI’s Legal Agent benchmark.
Despite its high performance, ML4 faces trade-offs in cost and speed compared to some specialized models. For instance, when compared to DeepSeek V4.1 Flash 1017 on the OpenRouter platform, Mistral Large 4 has a significantly higher input price ($1.36/M tokens vs. $0.049/M tokens) and a slower median throughput (27.0 tokens per second vs. 44.0 tokens per second). Furthermore, while ML4 is designed for high-reasoning tasks, some early user feedback suggests it may exhibit unexpected linguistic behaviors, such as producing Chinese characters during interactions in other languages like German. Mistral AI trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters to ensure independent operation under European law.