French AI laboratory Mistral has unveiled its new flagship model, Large 4 Le Chonk. The company describes it as its most powerful system to date and plans to develop several specialized versions based on this architecture. A preview version is currently accessible via API, with open weights scheduled for release by the end of October.
The model utilizes a Mixture-of-Experts (MoE) architecture, featuring 1 trillion parameters in total, with approximately 49 billion parameters activated per query. It is a multimodal system capable of processing text, images, complex documents, technical drawings, and charts. Beyond simple Q&A, the model can execute multi-step tasks, including code generation, financial analysis, and scientific calculations.
Mistral emphasizes the model's security capabilities, noting it achieved an 82% score in the Artificial Analysis Cyber Index and 93% in the Cybench test suite. Despite its open nature, the model includes safety protocols to reject dangerous instructions. The training process was conducted on 3,800 Nvidia Grace Blackwell accelerators in the company's European data centers.
The current version of Mistral Large 4 is undergoing further reinforcement learning to improve performance in the coming weeks and months. The company intends to use this model as a foundation for a new generation of industry-specific AI solutions.
