★ Reading this for free? Get 25 structured AI courses + per-chapter AI tutor — the first chapter of every course free, no card.Start free in 30 secondsOr own every course: $149 once

Ministral 3B: Mistral’s Smallest Model, Run Locally

Ministral 3 3B is a 3.4B-parameter language model with a 0.4B vision encoder, released by Mistral on 2 December 2025 under Apache 2.0. On Ollama it is ministral-3:3b, a 2.95 GB download. It is one of the smaller LLMs you can run locally and runs on almost any AI hardware.

Technical Specifications

DeveloperMistral AI
Released2 December 2025 (Ministral 3 family: 3B, 8B, 14B)
Parameters3.4B language model + 0.4B vision encoder
Layers / hidden size26 / 3,072
Attention32 query heads, 8 key-value heads (grouped-query)
Context256K (262,144 tokens)
VariantsBase, Instruct (published in FP8), Reasoning
InputsText and images
Languages (card)English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic and more
LicenceApache 2.0
Ollamaministral-3:3b, 2.95 GB (Q4_K_M)

From Mistral’s Hugging Face model card and config for Ministral-3-3B-Instruct-2512, and the Ollama registry (2 November 2026).

Which Ministral 3B?

There are two. Mistral announced the first Ministral 3B on 16 October 2024 alongside Ministral 8B. That 3B was available through Mistral’s API under a commercial licence; Mistral published weights only for Ministral 8B Instruct, for research use. You cannot download it.

The model you can run is Ministral 3 3B (December 2025). It comes as base, instruct and reasoning variants, all with image input and all under Apache 2.0, which allows commercial use. Everything below refers to it.

Published benchmark scores

All scores are from Mistral’s Ministral-3-3B-Instruct-2512 model card, where Mistral compares it with similar-sized models. They are Mistral’s evaluations, not ours. Mistral prints Arena Hard and MATH as fractions; they are shown here as percentages.

Instruct modelArena HardWildBenchMATH (maj@1)MM MTBench
Ministral 3 3B Instruct30.556.883.07.83
Qwen3-VL-4B-Instruct43.856.890.08.01
Qwen3-VL-2B-Instruct16.342.278.66.36
Gemma3-4B-Instruct31.849.175.95.23

Base model (pre-training)

  • MMLU, 5-shot: 70.7 (Qwen 3 4B Base 71.3; Gemma 3 4B Base 58.9)
  • MMLU Redux, 5-shot: 73.5
  • Multilingual MMLU: 65.2
  • MATH, CoT 2-shot: 60.1
  • AGIEval, 5-shot: 51.1
  • TriviaQA, 5-shot: 59.2

Reasoning variant

  • AIME 2025: 72.1 (Qwen3-VL-4B-Thinking 69.7)
  • AIME 2024: 77.5 (72.9)
  • GPQA Diamond: 53.4 (60.1)
  • LiveCodeBench: 54.8 (51.3)

The reasoning build is a separate model (Ministral-3-3B-Reasoning-2512); the Ollama ministral-3:3b tag is the instruct model.

Memory and speed

The 256K context is long, but memory grows with it. With 26 layers and 8 key-value heads the f16 KV cache is 2 × 26 × 8 × 128 × 2 bytes = 104 KB per token. Totals below are the 2.95 GB file plus that cache, arithmetic rather than measurement; allow about 1 GB more for the runtime.

ContextKV cacheWeights + cache
4,096 tokens0.44 GB~3.4 GB
32,768 tokens3.5 GB~6.4 GB
262,144 tokens (full 256K)27.9 GB~30.9 GB

Speed ceilings, not measurements. Generation of a dense model is capped by memory bandwidth divided by the file read per token. For the 2.95 GB ministral-3:3b that is ~92 tok/s on an RTX 4060 (272 GB/s), ~122 tok/s on an RTX 3060 12GB (360 GB/s), ~51 tok/s on an Apple M3 Pro (150 GB/s) and ~41 tok/s on a base Apple M4 (120 GB/s). This site has not run Ministral 3 3B. For scale, on our Apple M3 Pro 18 GB (Ollama 0.40.13, 2 November 2026) similar-sized models measured llama3.2:3b 55.6 tok/s, qwen3:4b 44.7 tok/s and phi4-mini 43.7 tok/s; Ministral’s file is larger than all three, so expect it to be somewhat slower on that machine.

Compact vs Traditional Performance

A 3B model is the right size when the job is narrow and the hardware is small: classification, extraction to JSON, short summaries, translation, captioning an image, or the first step of an agent that routes requests. Mistral’s card lists those same uses and highlights native function calling and JSON output.

It is the wrong size for long-form writing, hard reasoning or broad world knowledge, where 8B and 14B models do measurably better on Mistral’s own tables: base-model MMLU rises from 70.7 (3B) to 76.1 (8B) and 79.4 (14B). If you have 8 GB of VRAM, ministral-3:8b (6.02 GB) is the better default; the 3B is for machines where that does not fit, or where speed matters most.

Ollama Deployment Guide

Ollama Commands

Other builds: ollama pull ministral-3:3b-instruct-2512-q8_0 (4.47 GB) or ollama pull ministral-3:3b-instruct-2512-fp16 (7.71 GB).

Ministral 3B vs Other Small Models

Sizes from the Ollama registry, 2 November 2026; “RAM” is file + KV cache at 4K + about 1 GB, arithmetic. See also Ministral 8B, Phi-4 Mini and Llama 3.2 3B.

Frequently Asked Questions

Can I download the original Ministral 3B from October 2024?

No. Mistral offered that model through its API (ministral-3b-latest) under its commercial licence and published weights only for Ministral 8B, for research use. The open-weight 3B is Ministral 3 3B, released on 2 December 2025 under Apache 2.0; that is the model Ollama ships as ministral-3:3b.

How big is ministral-3:3b on Ollama?

2.95 GB. The default tag is the instruct model quantised to Q4_K_M. The registry also has ministral-3:3b-instruct-2512-q8_0 (4.47 GB) and ministral-3:3b-instruct-2512-fp16 (7.71 GB), plus the larger ministral-3:8b (6.02 GB) and ministral-3:14b (9.08 GB).

Can Ministral 3 3B read images?

Yes. Mistral describes it as a 3.4B language model plus a 0.4B vision encoder, and every Ministral 3 variant has image understanding. Mistral recommends keeping images close to a 1:1 aspect ratio.

How good is it compared with other small models?

On Mistral's own instruct table it ties Qwen3-VL-4B-Instruct on WildBench (56.8) and leads Gemma3-4B-Instruct (49.1), but trails Qwen3-VL-4B on Arena Hard (30.5 vs 43.8) and MATH (83.0 vs 90.0). These are Mistral's evaluations; no independent comparison is quoted here.

How fast is it?

This site has not timed it. As a bandwidth ceiling for the 2.95 GB file: ~92 tok/s on an RTX 4060, ~122 on an RTX 3060 12GB, ~51 on an Apple M3 Pro and ~41 on a base M4. Real runs land below the ceiling. On the M3 Pro, the smaller llama3.2:3b, phi4-mini and qwen3:4b files (2.0-2.5 GB) measured 44-56 tok/s, so expect Ministral's larger file to come in lower.

How much memory does the 256K context need?

The KV cache is about 104 KB per token at f16, so the full 262,144 tokens add 27.9 GB on top of the weights. At a 4K or 32K context it is 0.44 GB or 3.5 GB. Set num_ctx only as high as you need.

Sources

Reading now
Join the discussion

Build Real AI on Your Machine

RAG, agents, NLP, vision, and MLOps - chapters across 25 courses that take you from reading about AI to building AI.

🎯
AI Learning Path

Go from reading about AI to building with AI

25 structured courses. Hands-on projects. Runs on your machine. Start free.

Or own it for life — Lifetime $149 $599, pay once
Barsha Mohapatra

Written by

Barsha Mohapatra

Founder and Editor, Local AI Master · Founder, Little AI Master

Founder of Little AI Master and the editor behind every course and guide on Local AI Master. Years of tutoring students taught her where AI education actually fails.

📅 Published: September 28, 2025🔄 Last Updated: November 2, 2026✓ Manually Reviewed

Related Guides

Continue your local AI journey with these comprehensive guides

More on AI Models Directory
See the full AI Models Directory guide.
📚
Free · no account required

Grab the AI Starter Kit — career roadmap, cheat sheet, setup guide

No spam. Unsubscribe with one click.

🎯
AI Learning Path

Found your model? Now build something with it.

25 hands-on courses — RAG, agents, fine-tuning — all running locally. First chapter free, no card.

Or own it for life — Lifetime $149 $599, pay once
Free Tools & Calculators