Mistral Large 3 Review

Mistral's largest public model, released on 2 December 2025 and still the top of its lineup nine months later, which says something about the pace at the European end of the field. It is a sparse mixture-of-experts design with 675 billion total parameters and 41 billion active, paired with a 2.5 billion parameter vision encoder, so it handles text and images across a 256,000-token context.

The significant fact is the license: Apache 2.0, weights published openly, no revenue clause attached, which is a materially more permissive position than any comparably sized model from a US or Chinese lab. That makes it the default choice for teams who need a capable multimodal model they can host, audit and modify without a legal review first. On raw capability it trails the current frontier, but it was never priced or positioned to compete there. This grade is a provisional placeholder pending hands-on testing.