Ministral 3 3B

+ contribute

Ministral 3 3B is an open-weights language model from Mistral, released December 2025. Its published size is 3.85B parameters and a 262k-token context window. Licence: apache-2.0.

45.2±0.8#439 of 719 LLMsfrom 33 benchmark results across 8 sources

$ how it compares

The LLMs ranked either side of Ministral 3 3B, out of 828 with a score, then more from Mistral.

#	model	score	where	RAM	released
522	Qwen1.5 72B Chat Alibaba	45±1	local	~31GB	2024-02
523	speechless-starcoder2-15b uukuguy	45±1	local	~12GB	2024-03
524	Ministral 3 3B Mistral	45±1	api $0.10/M +local	~3.5GB	2025-12
525	Claude 2.0 Anthropic	45±1	comparison only		2023-07
526	OpenThinker 7B Open Thoughts	45±2	local	~6GB	2025-01
	more from Mistral
79	Mistral Large 4 (public preview) Mistral	71±2	api $2.06/M		2026-10
174	Mistral Medium 3.5 (128B) Mistral	65±1	api $3.00/M +local	~42GB	2026-04
180	Mistral Medium 3.1 (2508) Mistral	64±2	comparison only		2025-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #282 of 300 * · 1.9% of its score · 1 test

Beats GLM-4 9B Chat, DBRX Instruct, Olmo 3 7B Instruct here, though they rank higher overall. Trails Granite 4.0 H 350M, Qwen3 0.6B, Qwen2.5 1.5B Instruct here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
ContractBench success rate (33 observation-contract tool tasks)	6.1% ✎	▆··▆▄▂▄▂▆▆█▄ beats 0/23	very hard · avg 14%	▇ 1.9%	sharp

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #328 of 393 * · 3.9% of its score · 2 tests

Beats OpenThinker3 7B, Gemma 4 E2B, Code Llama 70B Instruct here, though they rank higher overall. Trails LFM2 350M, Gemma 3 1B, Flan-T5 XXL here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena-Hard (version/judge unspecified) lab	30.5% ✎	▃▃▂▅▃▆▅█▅▅▂▃ beats 3/30	medium · avg 48%	▇ 2.7%	at its level · small run
IFEval (aggregation unspecified)	72.5% ✎	▄···▄█▄····▄ beats 2/5	easy · avg 79%	▇ 1.2%	at its level

math▁▂▃▅▆▅▄█▄▁▁▁ #269 of 398 · 28% of its score · 4 tests

Beats Molmo 2 8B, DeepSeek-V2.5, DeepSeek-Coder-V2 Instruct (0724) here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Mixtral 8x7B v0.1 (base), InternVL2-4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MATH (subset unspecified) +3	70.2% ✎	█▄··▄█▄····▄ beats 0/7	easy · avg 81%	▇▇▇▇▇ 14%	at its level · several runs
GSM8K	79.0% ✎	█····█··▄·▄█ beats 1/7	very easy · avg 86%	▇▇▇▇ 9.4%	sharp
MGSM (Marco evaluation)	47.0% ✎	▄·▄·▄▄▄·▄··█ beats 0/7	easy · avg 70%	▇ 3.0%	at its level
PolyMath (Marco evaluation)	16.3% ✎	▄··▄██··▄··▄ beats 0/7	hard · avg 30%	▇ 2.5%	typical

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #130 of 147 · 11% of its score · 9 tests

Beats Gemma 4 E2B here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
CharXiv descriptive questions	76.6% ✎	█·▄·······▄█ beats 5/5	easy · avg 78%	▇ 3.4%	at its level
MMStar	53.1% ✎	▄▄▄█·······▄ beats 4/5	medium · avg 62%	▇ 2.5%	at its level
MMBench-EN dev v1.1	69.2% ✎	█·····▃··▃·▃ beats 1/5	easy · avg 79%	▇ 1.4%	blunt · at its level
HallusionBench	65.2% ✎	▄▄·▄·····█·▄ beats 3/5	medium · avg 65%	▇ 1.2%	blunt · at its level
RealWorldQA	58.3% ✎	▄···█··▄▄··▄ beats 2/5	easy · avg 69%	▇ 0.6%	blunt · at its level
CountBench	73.7% ✎	▄··▄··█··▄·▄ beats 3/5	very easy · avg 86%	▇ 0.6%	noisy
OCRBench	73.5% ✎	█··███·█···█ beats 2/5	easy · avg 81%	▇ 0.4%	blunt
AI2D test	74.1% ✎	█···▄▄···▄·▄ beats 2/5	easy · avg 82%	▇ 0.3%	blunt · at its level

1 more test

test	result	vs the others in its run	difficulty	counts	why
ChartQA	79.1% ✎	▄······▄··██ beats 3/5	easy · avg 80%	▇ 0.2%	blunt · noisy

other tests13 not in a task · 55% of its score

test	result	vs the others in its run	difficulty	counts	why
MMLU +1	69.8% ✎	▄···▄▄·▄█··█ beats 1/7	easy · avg 75%	▇▇▇▇▇ 11%	at its level · several runs
Medmarks-V mean win rate (19 verifiable medical benchmarks, 71 configurations)	0.3927 ✎	▁▄▃▃▇▄█▃▅▁▁▁ beats 4/30	avg 0.423	▇▇▇▇ 10%	sharp
MMLU-Pro +1	49.5% ✎	▄····█·▄█·▄▄ beats 2/7	medium · avg 56%	▇▇▇ 8.5%	at its level · several runs
GPQA Diamond	31.6% ✎	▄···█·▄·█·▄▄ beats 1/7	hard · avg 38%	▇▇ 6.0%	sharp
Medmarks-V weighted mean win rate (61 models, 71 configurations)	39.3% ✎	▂▃▅▂▆▃█▅█▅▃▆ beats 4/34	medium · avg 42%	▇▇ 4.4%	blunt · at its level
MMLU-ProX-Lite (Marco)	26.7% ✎	█··▄··█··█·▄ beats 0/7	medium · avg 43%	▇▇ 4.1%	at its level
GlobalMMLU (Marco evaluation)	38.4% ✎	▃··▃▃··▅···█ beats 0/7	medium · avg 57%	▇ 2.9%	at its level
MMMLU (Marco evaluation)	39.4% ✎	▄···▄·█·▄·▄█ beats 0/7	medium · avg 60%	▇ 2.6%	at its level

5 more tests

test	result	vs the others in its run	difficulty	counts	why
AGIEval (Marco suite unspecified)	44.7% ✎	▄··▄█▄··▄··█ beats 3/7	medium · avg 47%	▇ 2.1%	blunt · at its level
Multi-IF (aggregation unspecified)	47.0% ✎	▄·▄··█▄····▄ beats 3/5	medium · avg 58%	▇ 1.3%	blunt · at its level
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy	69.8% ✎	▃·▁▁▂▁▂▅█▇▄▁ beats 50/100	medium · avg 68%	▇ 1.1%	at its level · noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy	81.5% ✎	▁▁▁▁▁▁▁▂▄█▄▃ beats 72/146	easy · avg 83%	▇ 0.9%	at its level · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	75.8% ✎	▃▂▁▂▅█▂▃▄▃▁▁ beats 0/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $0.10/M · $0.1 in / $0.1 out per 1M tokens · 262k context · docs.mistral.ai ↗ · price sourceUSD per 1M text tokens. Mistral standard synchronous API; model ministral-3b-2512. Uncached input; output includes generated reasoning where billed. Batch, cache storage, tools and other modalities excluded.

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Ministral 3 3B · needs ~2.1 GB · from huggingface.co/mistralai
chat in the terminal
llama cli -hf mistralai/Ministral-3-3B-Instruct-2512-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf mistralai/Ministral-3-3B-Instruct-2512-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Ministral 3 3B · needs ~2.1 GB · from huggingface.co/mistralai
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Ministral 3 3B · needs ~2.81 GB · from ollama.com
chat in the terminal
ollama run ministral-3:3b

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Ministral 3 3B · needs ~1.9 GB, Mac only · from huggingface.co/Oscilla
chat in the terminal
mlx_lm.chat --model Oscilla/Ministral-3-3B-Instruct-2512-mlx-4Bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model Oscilla/Ministral-3-3B-Instruct-2512-mlx-4Bit

Weights and documentation: huggingface.co/mistralai ↗

$ how the score was worked out · Marco-MoE: Open Multilingual… 51%, CohereLabs/North-Micro-Vision… 19%, Medmarks: A Comprehensive Ope… 14% and 5 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcycling	11	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 51%	CC BY 4.0
CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)	13	▇▇▇▇▇▇ 19%	Apache-2.0
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models	2	▇▇▇▇ 14%	CC BY 4.0
mistralai/Ministral-3-14B-Instruct-2512 model card · lab-reported	2	▇▇ 7%	Apache-2.0
mistralai/Ministral-3-8B-Instruct-2512 model card · lab-reported	1	▇ 4%	Apache-2.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇ 2%	MIT
ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)	1	▇ 2%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 45.2