Mistral previews one-trillion-parameter Large 4 model

Mistral has introduced a public preview of Large 4, a one-trillion-parameter open-weight AI model the French company says is designed for coding, cybersecurity and business applications. The model is available through Mistral’s API, while the company says its downloadable weights are planned for release on October 27, allowing users to run it on their own systems. Mistral says Large 4 can process text and images and uses 49 billion of its one trillion parameters for any task. The company says it trained the model from scratch over two months on about 4,000 Nvidia Grace Blackwell GPUs in its European data centres. The training figure comes from Mistral; another account gave a lower count, and the available information does not clarify the difference. The preview is not the final version. Mistral says reinforcement-learning work is still underway and that benchmark results are preliminary and may change before the weights are released. During testing, selected developers, cybersecurity firms and government agencies are to receive a version with fewer restrictions and expanded cyber capabilities, according to the company.