Model releases · 11 August 2026

Ministral 3 8B: a small Mistral with vision, under Apache 2.0

Mistral's compact line carries a permissive licence, image input and a quarter-million-token context — an unusual combination at eight billion parameters.

At a glance

Released by	Mistral AI
Current checkpoint	Ministral-3-8B-Instruct-2512
Parameters	8B
Architecture	`Mistral3ForConditionalGeneration` (`mistral3`)
Layers / hidden size	34 / 4096
Context length	262,144 tokens
Vocabulary	131,072 tokens
Modalities	Text and image
Licence	Apache 2.0
Weight size	20.88 GB
Variants	Instruct and Reasoning, plus a BF16 build
Hosted at	`mistralai/Ministral-3-8B-Instruct-2512` on Hugging Face

What it is

Ministral 3 8B is the balanced member of Mistral's compact family — the line the company positions as tiny language models with vision. Mistral publishes it under Apache 2.0, explicitly noting the licence permits commercial and non-commercial use and modification.

Three things worth noting at this size

Vision is included. The architecture is Mistral3ForConditionalGeneration, not a text-only causal model. Image input at 8B, under Apache 2.0, is not common.
262K context. A quarter of a million tokens on a model this size puts it alongside releases several times larger.
A separate reasoning variant. Mistral publishes Ministral-3-8B-Reasoning-2512 alongside the instruct model rather than folding a thinking mode into one checkpoint.

Where it sits

Mistral's wider catalogue has moved upward in size — Mistral Small 4 at 119B, Mistral Medium 3.5 at 128B, Mistral Large 3 at 675B. Ministral 3 is the part of the range aimed at deployment on modest hardware, and the 8B is its middle rung.

The company also published Shieldstral 1.0 3B in the same period, a safety-classification model, also Apache 2.0.

Runtime support

Runtime	Supported	Notes
transformers	Yes	Reference implementation from Mistral.
llama.cpp	Partial	Text support for the Mistral 3 line exists; vision requires a separate projector and lags behind.
MLX	Partial	Community conversions; nothing official from Mistral.
Vision input	Yes	Built into the architecture rather than added by a downstream projector.

What we don't know yet

Mistral does not publish official quantised builds for this checkpoint.
How well the 262K context holds up in practice at this parameter count is not characterised.
The trade-off between the Instruct and Reasoning variants is not quantified in the model card.

Sources

Frequently asked questions

What licence is Ministral 3 8B under?

Apache 2.0, permitting commercial and non-commercial use and modification.

Does Ministral 3 8B support images?

Yes. Vision is part of the architecture rather than an add-on, which is uncommon for an Apache-licensed model at this size.

What context length does Ministral 3 8B support?

262,144 tokens.

Is there a reasoning version?

Yes. Mistral publishes Ministral-3-8B-Reasoning-2512 as a separate checkpoint rather than combining modes in one model.

How large is the download?

20.88 GB for the published safetensors weights. Mistral does not publish official quantised builds for this checkpoint.

OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.

Download on the App Store Muse Glimmer 30B Motif 3

Published 11 August 2026