Ministral 3

What it is

Ministral 3 is a family of three small dense models from Mistral AI, with sizes labelled 3B, 8B and 14B. All three were released on 2 December 2025 alongside Mistral Large 3, so the small models arrived as part of the same launch rather than as a later follow-up^[1].

Every size is published in three variants — base, instruct and reasoning — and all of them can understand images as well as text^[1]. The weights are released under the Apache 2.0 licence^[1], the permissive open-source licence that allows reuse and modification with few conditions. In Mistral's own documentation the family is filed under version 25.12^[2].

“Small” and “dense” describe two different things. Small refers to the parameter count: these models are meant to run on a single device rather than on rows of data-centre accelerators. Dense means that every parameter is used to process every token, in contrast with mixture-of-experts designs that activate only part of the network at a time.

The three sizes

Mistral's models overview gives each size a short role. The 14B is described as offering best-in-class text and vision capabilities — the largest of the three^[2]. The 8B is described as a powerful and efficient model, again with best-in-class text and vision capabilities: the middle option, weighing quality against the cost of running it^[2]. The 3B is described as tiny and efficient, with the same best-in-class text and vision claim^[2]. The repeated wording matters less than the shape of the offer: all three handle vision, so a developer choosing between them is trading quality against footprint, not deciding which modalities to give up.

Reasoning performance

Mistral reports that the reasoning variant of the 14B model scored 85% on AIME 2025, the 2025 edition of a mathematics competition written for strong high-school students^[1]. A benchmark of this kind is the usual way model makers signal how well a reasoning mode — one that works through intermediate steps before giving an answer — copes with multi-step mathematics. It is one number on one benchmark, and it says nothing directly about how the model performs on other tasks.

Where it runs

Mistral says the family is optimised for deployment on NVIDIA DGX Spark, on RTX PCs and on Jetson devices^[1]. Those are three quite different classes of hardware: a compact desktop AI machine, consumer graphics cards, and embedded modules for robots and edge devices. The common thread is local execution — the models are meant to be usable on hardware that a person or a small team owns, without a round trip to a hosted API.

Place in the lineup

In Mistral's API lineup, the new models replace the earlier Ministral 3B and 8B from 2024, along with Pixtral 12B and Mistral NeMo^[2]. That consolidation is a useful snapshot of how the small-model tier has changed: separate families for vision (Pixtral) and for compact general-purpose work (NeMo, Ministral) fold into one line that covers text, images and a reasoning mode at three sizes. For anyone building on Mistral's API the practical consequence is fewer model names to choose between; for anyone running models locally, it is an openly licensed family that covers a wider range of hardware under one naming scheme.