Mistral AI released Mistral 3 as a family spanning local devices and data centers. The smaller Ministral models come in three sizes, each with a starting model for customization, an instruction-following version, and a reasoning version designed to spend longer on answers. All support image understanding. Large 3 serves the larger deployment tier, with base and instruction-following versions available at launch. The December 2025 announcement places the whole family under Apache 2.0, giving developers multiple sizes and training variants they can deploy and adapt themselves.
Terms in this piece · Glossary
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
