Ministral 8B

A powerful and efficient model in Mistral's Ministral 3 family, offering best-in-class text and vision capability at its size. It is built for edge deployment and performs across diverse hardware, fitting in 12GB of VRAM at FP8 and less when further quantised, while carrying a 256K context and image understanding across dozens of languages. It does not reason -- the reasoning capability ships as a separate checkpoint rather than a runtime toggle.

The 8B sits in the middle of a three-size edge family - 3B, 8B and 14B - released as the small end of the Mistral 3 generation. Mistral gave every size three separate checkpoints rather than one switchable model: a base, an instruct and a reasoning variant, each with image understanding, all under Apache 2.0. That is the structural fact worth knowing before you pick this one, because it means the reasoning capability is a different download rather than a runtime flag.

Mistral's claim for the family is a ratio rather than a peak: it says the Ministral models offer the best cost-to-performance of any open model, and it argues the case in a way most small-model launches do not. Its point is that in real deployments the tokens generated matter as much as the parameter count, and it reports the instruct variants matching or beating comparable models while often emitting an order of magnitude fewer tokens to get there. The chart it published makes that concrete for this size: the 8B instruct model lands around 51 on GPQA Diamond at roughly 1,500 output tokens, in the same accuracy band as a comparable 8B model that spends more than ten times as many.

The lab is explicit about the division of labour inside the family. For settings where accuracy is the only concern, it points at the reasoning variants, which think longer to reach what it calls state-of-the-art accuracy for their weight class - it cites 85% on AIME 2025 for the 14B. This instruct checkpoint is the opposite trade: answer fast, answer short, stay cheap.

The deployment story is where the size pays. Mistral says the whole Mistral 3 family was trained on NVIDIA Hopper hardware and that it worked with NVIDIA on optimised deployments for desktop AI machines, RTX PCs and laptops, and embedded Jetson boards - the point being that the 8B is meant to run on the device rather than in a datacentre. It carries native multilingual coverage across more than forty languages and image understanding at the same size. What it is not is a frontier model: Mistral positions it for edge inference, classification, routing and on-device assistants, and points at its far larger models for anything that needs frontier reasoning.