Mistral AI has unveiled Mistral Large 4, a multimodal model featuring 1.05 trillion parameters. The model utilizes a Mixture of Experts architecture, with 49 billion active parameters per request, and includes a 1.6 billion parameter visual encoder. It supports a context window of up to 1 million tokens and is designed for tasks such as programming, cybersecurity, finance, and chip design.
The company trained the model over two months using approximately 4,000 NVIDIA Grace Blackwell GPUs within its European data centers. Mistral Large 4 is currently available in public preview via API and the company's cloud infrastructure. A full release of the open weights is scheduled for October 27, with early access currently provided to select cybersecurity and government partners.
Mistral AI emphasizes that the open-weights approach allows developers to deploy the model independently without sharing data with external providers. This release follows a recent funding round that valued the company at over $24 billion, with participation from investors including Samsung Electronics. The company is currently testing the model for safety, specifically regarding autonomous behavior.
The launch occurs amidst intensifying competition in the open-weights sector, where European developers are challenging both closed-source systems and emerging Chinese models. Mistral aims to bridge the gap by providing high-performance capabilities while maintaining user control over infrastructure and model weights.
