Ministral 3 8B

Ministral 3 8B is the balanced member of the Ministral 3 family that Mistral AI released on 2 December 2025. It splits into two blocks the vendor describes separately: an 8.4-billion-parameter language model and a 0.4-billion-parameter vision encoder, 8.9 billion weights in total. The instruct version ships in FP8 precision, and that is the decisive number for its audience — at FP8 the model fits inside 12 GB of video memory, which is what an ordinary consumer graphics card offers. A multimodal assistant with a 256,000-token context therefore runs without a server, a subscription or a network connection. Mistral publishes base, instruct and reasoning variants of this size under Apache 2.0. The lab's own framing for the family is cost-to-performance rather than raw benchmark position: the instruct models are supposed to match comparable competitors while producing an order of magnitude fewer tokens on the way to the answer, which matters more in local deployment, where every generated token is time on the user's own hardware. For those who prefer not to host it, the model is served on the vendor's platform under the identifier ministral-8b-2512 at 0.15 US dollars per million tokens, the same rate for input and output. Mistral recommends running it at a temperature below 0.1 for everyday use and feeding it images close to a 1:1 aspect ratio, since the vision encoder degrades on very wide or very narrow crops.

▸Videos

No videos yet.

▸Reviews