Mistral AI Unveils 1-Trillion-Parameter Large 4 Model to Challenge US and Chinese AI Rivals

Mistral Large 4's introduction aims to establish a "third way in AI," according to French President Emmanuel Macron, by offering an alternative to closed American systems and open-weight models from China. The company emphasizes its multimodal capabilities and targets demanding enterprise applications across cybersecurity, coding, finance, manufacturing, geospatial analysis, and product design.

Mistral CEO Artur Mensch said at a launch event in Abu Dhabi that the new model was "above the Chinese models on certain aspects, including cyber." The company plans to release the model's weights to developers, cybersecurity leaders, government authorities and other selected users after an initial safety-testing period, intending for its open-source weights to be used for defence rather than malicious attacks.

The company stated that Large 4 was trained on its own European infrastructure using 3,800 NVIDIA Grace Blackwell GPUs. Pierre Stock, Mistral's vice president of science, told TechCrunch that the training used about 4,000 GPUs. Stock also said the computing resources used were two to three times lower than those used by Mistral's Chinese competitors and significantly lower than those used by closed-source competitors, a comparison based on Mistral's own assessment.