whatmodel.sh / Ministral 3 14B

Ministral 3 14B

+ contribute

Ministral 3 14B is an open-weights language model from Mistral, released December 2025. Its published size is 13.9B parameters and a 262k-token context window. Licence: apache-2.0.

55.7±1.1#253 of 719 LLMsfrom 17 benchmark results across 7 sources

$ how it compares

The LLMs ranked either side of Ministral 3 14B, out of 828 with a score, then more from Mistral.

#	model	score	where	RAM	released
302	Qwen3 8B Alibaba	56±1	api $0.31/M +local	~6.1GB	2025-04
303	DeepSeek R1 Distill Qwen 14B DeepSeek	56±1	local	~10GB	2025-01
304	Ministral 3 14B Mistral	56±1	api $0.20/M +local	~9.1GB	2025-12
305	Nemotron Nano 9B v2 NVIDIA	56±1	api $0.10/M +local	~7.6GB	2025-08
306	Magistral Medium 1.0 (2506) Mistral	56±1	comparison only		2025-06
	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▁▁▃▆█▆▅▅▅▄▃▁ #204 of 300 · 12% of its score · 3 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 (Jan '25) here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Mellum2 12B-A2.5B 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)	28.3% ✎	▆··▆▄▃▄▂▆▆█▄ beats 8/23	very hard · avg 14%	▇▇▇ 7.6%	sharp
BFCL V4 (December 2025 archive)	38.8% ✎	▁▁▁▂▇▆▆▆█▂▄▂ beats 51/111	medium · avg 36%	▇ 2.9%	at its level
BFCL v3	52.7% ✎	▁▁▁▁▁▃▅█▆▅▄▁ beats 56/170	medium · avg 56%	▇ 1.4%	blunt · at its level

coding▁▁▁▃▆██▇▆▄▃▁ #289 of 458 * · 11% of its score · 2 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails DeepSeek LLM 67B Chat, WizardLM 70B V1.0, Yi 1.5 9B Chat here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench v6 (window unspecified)	42.4% ✎	▂▃▄▄▅██▅▄▇▃▆ beats 10/65	hard · avg 34%	▇▇▇▇ 8.9%	at its level
MultiPL-E	71.5% ✎	▃··▄·▅▅▇▄█▃▄ beats 19/31	medium · avg 55%	▇ 1.9%	at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #276 of 393 * · 6.2% of its score · 2 tests

Beats Ministral 3 14B Reasoning 2512 here, though they rank higher overall. Trails Flan-T5 XXL, EXAONE 3.5 2.4B Instruct, LFM2 1.2B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena-Hard (version/judge unspecified) lab	55.1% ✎	▃▃▂▅▃▆▅█▅▅▂▃ beats 10/30	medium · avg 48%	▇▇ 4.8%	at its level · small run
IFEval (aggregation unspecified)	67.3% ✎	▁▁▃▁▃▄▄▆█▃▂▁ beats 30/141	easy · avg 76%	▇ 1.4%	typical

math▁▂▃▅▆▅▄█▄▁▁▁ #197 of 398 * · 10% of its score · 1 test

Beats MiniMax-M2.1, Qwen3 235B-A22B (Apr 2025, hybrid), DeepSeek V3 (Dec '24) here, though they rank higher overall. Trails Phi-4-mini-reasoning 3.8B, DeepSeek R1 Distill Qwen 7B, OpenThinker2 7B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MATH (subset unspecified) lab +1	90.4% ✎	▂▄▄▃█▇▅▆▄▃▄▁ beats 66/80	easy · avg 83%	▇▇▇▇ 10%	small run · several runs

other tests8 not in a task · 61% of its score

test	result	vs the others in its run	difficulty	counts	why
Medmarks-V mean win rate (19 verifiable medical benchmarks, 71 configurations)	0.4760 ✎	▁▄▃▃▇▄█▃▅▁▁▁ beats 13/30	avg 0.423	▇▇▇▇▇▇▇ 17%	sharp
GPQA Diamond	58.6% ✎	▁▁▁▁▂▄▅▅███▄ beats 107/310	hard · avg 47%	▇▇▇▇▇▇ 16%	sharp · at its level
Medmarks-V weighted mean win rate (61 models, 71 configurations)	47.6% ✎	▂▃▅▂▆▃█▅█▅▃▆ beats 12/34	medium · avg 42%	▇▇▇ 8.2%	blunt · at its level
EuroEval Life in the UK (English knowledge MC), validation split, 5-shot, accuracy	92.7% ✎	▂·▂▂·▂·▃▃·█▃ beats 8/14	easy · avg 80%	▇▇▇ 7.3%	typical
MMLU-Redux	85.9% ✎	▁▁▃▁▂▃▄▅▇▅█▄ beats 21/63	easy · avg 78%	▇▇▇ 7.0%	typical
EuroEval HellaSwag (English), validation split, 5-shot generative MC, accuracy	79.3% ✎	▂··▂▂·▂▆▆█·▄ beats 10/15	medium · avg 63%	▇▇ 4.5%	at its level
GSM-Plus	86.6% ✎	▁··▁▁▁·▃▁▄█▁ beats 9/16	easy · avg 82%	▇ 1.4%	blunt
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	80.6% ✎	▁▃▁▂▅█▂▃▄▃▁▁ beats 5/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $0.20/M · $0.2 in / $0.2 out per 1M tokens · 262k context · docs.mistral.ai ↗ · price sourceUSD per 1M text tokens. Mistral standard synchronous API; model ministral-14b-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 14B · needs ~8.2 GB · from huggingface.co/mistralai
chat in the terminal
llama cli -hf mistralai/Ministral-3-14B-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-3-14B-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 14B · needs ~8.2 GB · from huggingface.co/mistralai
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/mistralai/Ministral-3-14B-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 14B · needs ~8.84 GB · from ollama.com
chat in the terminal
ollama run ministral-3:14b

MLX (Mac)

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Ministral 3 14B · needs ~7.2 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/mistralai_Ministral-3-14B-Instruct-2512-MLX-MXFP4
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/mistralai_Ministral-3-14B-Instruct-2512-MLX-MXFP4

Weights and documentation: huggingface.co/mistralai ↗

$ how the score was worked out · JetBrains/Mellum2-12B-A2.5B-I… 41%, Medmarks: A Comprehensive Ope… 25%, EuroEval leaderboards (result… 12% and 4 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
JetBrains/Mellum2-12B-A2.5B-Instruct model card	8	▇▇▇▇▇▇▇▇▇▇▇▇ 41%	Apache-2.0
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models	2	▇▇▇▇▇▇▇ 25%	CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇▇▇▇ 12%	MIT
mistralai/Ministral-3-14B-Instruct-2512 model card · lab-reported	2	▇▇▇ 10%	Apache-2.0
ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)	1	▇▇ 8%	CC BY 4.0
mistralai/Ministral-3-8B-Instruct-2512 model card · lab-reported	1	▇ 5%	Apache-2.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 1%
weighted average		= 55.7

$ quantization evidence · 8 measurements

How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent.

precision	test	full → quantized	kept
FP8 W8A8 dynamic (weights per-channel, activations per-token; llm-compressor)	GPQA Diamond (lighteval, vLLM, temperature 0.15)	0.586 → 0.662	112.9%	source ↗
FP8 W8A8 dynamic (weights per-channel, activations per-token; llm-compressor)	Reasoning average of 3 tasks (AIME 2025, GPQA Diamond, and a third task labelled "Math-lvl-5" in the table but run with lighteval `math_500` per the card commands)	0.613 → 0.619	100.9%	source ↗
FP8 W8A8 dynamic (weights per-channel, activations per-token; llm-compressor)	IFEval 0-shot (lm-evaluation-harness, chat template, vLLM)	0.773 → 0.769	99.4%	source ↗
FP8 W8A8 dynamic (weights per-channel, activations per-token; llm-compressor)	AIME 2025 (lighteval, vLLM, temperature 0.15)	0.367 → 0.3	81.8%	source ↗
NVFP4 (FP4 weights and activations, llm-compressor NVFP4 scheme)	GPQA Diamond (lighteval, vLLM, temperature 0.15)	0.586 → 0.609	104.0%	source ↗
NVFP4 (FP4 weights and activations, llm-compressor NVFP4 scheme)	Reasoning average of 3 tasks (AIME 2025, GPQA Diamond, and a third task labelled "Math-lvl-5" in the table but run with lighteval `math_500` per the card commands)	0.613 → 0.598	97.5%	source ↗
NVFP4 (FP4 weights and activations, llm-compressor NVFP4 scheme)	AIME 2025 (lighteval, vLLM, temperature 0.15)	0.367 → 0.325	88.6%	source ↗
NVFP4 (FP4 weights and activations, llm-compressor NVFP4 scheme)	IFEval 0-shot (lm-evaluation-harness, chat template, vLLM)	0.773 → 0.635	82.2%	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json