Paris, October 9, 2026
Mistral AI SAS, a leading French artificial intelligence company, has unveiled its most sophisticated large language model (LLM) yet, Mistral Large 4, through a public preview on its cloud platform. The model, boasting a mixture of experts architecture with 1 trillion parameters, is designed for hardware efficiency by activating only 49 billion parameters when processing user prompts. Mistral Large 4 can answer questions in over 160 languages and achieved a top-five score on the AA Cyber Index, a benchmark that assesses LLMs’ ability to detect and fix software vulnerabilities. In patching open-source projects, Mistral Large 4 scored 82%, outperforming many open-source competitors. Compared to GPT-6 Astra, Mistral Large 4 lags behind on popular coding benchmarks but outperforms several leading open-source LLMs such as Qwen3.8 Max and DeepSeek V4 Pro on the AutomationBench and AA-Briefcase knowledge work benchmarks. These latter benchmarks include complex tasks that would typically take a human weeks to complete. Mistral Large 4 was trained on 3,800 Grace Blackwell chips, each combining two Nvidia Blackwell graphics cards with one central processing unit. The training process involved tens of thousands of rollouts, trial-and-error exercises used to refine the model. These rollouts produced 33 billion tokens per day, with nearly half used for training. Mistral plans to release the model’s weights later this month and expects the workflow to produce larger, more capable versions of the LLM in the coming months. The company also intends to build a series of models optimized for specific use cases using Mistral Large 4 as a foundation. “The launch of Mistral Large 4 marks a significant step forward in the development of open-source AI models, emphasizing both performance and efficiency,” said Dr. Alexis Boutin, a researcher at Mistral AI. With Mistral Large 4, the French AI company joins the race to create powerful and efficient LLMs that can compete with frontier models like Astra while outperforming many open-source alternatives on specific tasks. The future promises larger and more capable versions of Mistral Large 4 and a series of models optimized for specific use cases, signaling a new era in the world of artificial intelligence.
