Qwen3 4B

+ contribute

Qwen3 4B is an open-weights language model from Alibaba, released April 2025. Its published size is 4.02B parameters and a 33k-token context window. Licence: apache-2.0.

Catalogue identity notes

Comparison model from the Epoch AI internal-evaluation CSVs (Epoch release date 2025-04-29). Original April 2025 hybrid-thinking checkpoint Qwen/Qwen3-4B; distinct from qwen3-4b-instruct-2507 / -thinking-2507. Default (thinking) runs used where present; the chess result is the only published run ('qwen3-4b_none'). 4,022,468,096 parameters. ollama qwen3:4b now points to the Thinking-2507 build, so the original tags are pinned explicitly.

50.9±0.6#327 of 719 LLMsfrom 81 benchmark results across 14 sources

$ how it compares

The LLMs ranked either side of Qwen3 4B, out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
393	Llama 3.1 Nemotron 51B Instruct NVIDIA	51±2	local	~34GB	2024-09
394	Ministral 3 3B Reasoning 2512 Mistral	51±1	local	~3.5GB	2025-12
395	Qwen3 4B Alibaba	51±1	local	~3.7GB	2025-04
396	Yi-Large 01.AI	51±1	comparison only		2024-05
397	Llama-3-Instruct-8B-SimPO-ExPO Chujie Zheng (ExPO)	51±1	local	~6.6GB	2024-05
	more from Alibaba
18	Qwen3.8 Max (2.4T-A95B) Alibaba	78±1	api $3.00/M +local	~380GB	2026-08
34	Qwen3.8 Max (0902) Alibaba	76±1	api $3.00/M		2026-09
37	Qwen3.8-Flash-Next Alibaba	75±1	local	~59GB	2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #214 of 300 · 16% of its score · 8 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, Granite 4.1 3B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
τ²-bench Telecom lab +1	17.5% ✎	▁·▅█▆▅▂▃▃▂▃▁ beats 11/89	very hard · avg 14%	▇▇ 4.9%	sharp · small run · several runs
τ-bench Retail lab +1	33.9% ✎	▁····▃·▂▇█▂▆ beats 3/28	hard · avg 31%	▇ 2.8%	at its level · small run · several runs
τ²-bench Retail lab +1	38.6% ✎	▁····▃▁▃▆█▃▁ beats 2/44	hard · avg 30%	▇ 2.3%	at its level · small run · several runs
τ-bench Airline lab +1	32.0% ✎	▁▁▁·▅▄▁█▇▁▁▇ beats 3/28	hard · avg 28%	▇ 1.9%	at its level · small run · several runs
τ²-bench Airline lab +1	28.0% ✎	▁▃·▂▁▃▂▁█▅▂▃ beats 1/42	hard · avg 18%	▇ 1.6%	small run · several runs
BFCL v3 +2	63.4% ✎	▁▁▁▁▁▃▅█▅▄▁▁ beats 91/114	medium · avg 52%	▇ 1.5%	blunt · at its level · several runs
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)	29.9% ✎	▂·▃▁··▇▇█▂·▁ beats 2/25	hard · avg 29%	▇ 0.7%	blunt · at its level
BFCL (version not stated)	95.5% ✎	▄▆▄▂·█▆·▄··▃ beats 17/17	very easy · avg 86%	▇ 0.5%	blunt · small run

coding▁▁▁▃▆██▇▆▄▃▁ #292 of 458 · 7.7% of its score · 4 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
Aider polyglot (225 exercises, two attempts) lab	13.8% ✎	▂▁▁▃▆▃▄▆█▃▄▂ beats 13/71	very hard · avg 8.9%	▇ 3.3%	sharp · small run
LiveCodeBench v6 (Feb–May 2025) lab +1	48.4% ✎	▁▁▃▃█▅▄▇▅▅▃█ beats 4/38	medium · avg 43%	▇ 2.5%	at its level · small run · several runs
MultiPL-E lab	66.6% ✎	▃··▄·▅▅▇▄█▃▄ beats 15/31	easy · avg 66%	▇ 1.1%	at its level · small run
LiveCodeBench v4	52.9% ✎	█·█········█ beats 2/2	medium · avg 47%	▇ 0.9%	at its level · small run

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #244 of 393 · 9.5% of its score · 5 tests

Beats Mistral Small 4 (119B-A6.5B), Ministral 3 14B Reasoning 2512, Qwen2.5 Coder 32B Instruct here, though they rank higher overall. Trails Flan-T5 XXL, Granite 4.0 H 1B, LFM2 2.6B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena-Hard v2 aggregate lab +1	13.7% ✎	▃▁··▄▃▁▃█▁▁▇ beats 0/22	hard · avg 20%	▇ 3.1%	sharp · small run · several runs
Arena-Hard v2.0 hard prompts (style control, Gemini-2.5 judge)	15.0% ✎	▂·▃▅▅·▃█▆▂▂▂ beats 4/22	very hard · avg 12%	▇ 2.3%	sharp
IFEval (aggregation unspecified) lab +2	81.9% ✎	▁▁▃▁▂▄▄▆█▃▂▁ beats 76/141	easy · avg 80%	▇ 2.2%	small run · several runs
IFEval mean of four measures	85.6% ✎	█··▄▄·▄····▄ beats 5/5	easy · avg 85%	▇ 1.3%	typical
Arena-Hard v2.0 creative writing (GPT-4.1 + Gemini-2.5 judges)	13.2% ✎	▆▃▂▆▂▂▃█▂▃·▂ beats 4/22	hard · avg 19%	▇ 0.6%	sharp · noisy

math▁▂▃▅▆▅▄█▄▁▁▁ #214 of 398 · 15% of its score · 5 tests

Beats DeepSeek V3 (Dec '24), GPT-4o (2024-08-06), Command A (03-2025) here, though they rank higher overall. Trails Olmo 3 7B Instruct, DeepSeek R1 Distill Llama 8B, Phi-4-mini-reasoning 3.8B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2025 lab +2	65.6% ✎	▁·▁▁▁▂▅▆█▇▅▁ beats 56/133	medium · avg 59%	▇▇▇ 7.1%	sharp · at its level · small run · several runs
HMMT 2025 (edition unspecified) lab +1	42.1% ✎	▃█▇▅▇▅▁·▁··▁ beats 5/32	hard · avg 34%	▇▇ 4.5%	sharp · at its level · small run · several runs
PolyMATH lab +1	40.0% ✎	▁···▃▁·▇█▄▄▄ beats 4/23	hard · avg 34%	▇ 1.9%	at its level · small run · several runs
MATH-500	85.6% ✎	▃····▃·█···▃ beats 5/5	easy · avg 81%	▇ 1.1%	typical
GSM8K loose substring accuracy	68.5% ✎	█·█··███···█ beats 0/5	easy · avg 84%	▇ 0.1%	blunt

other tests30 not in a task · 52% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond +2	52.3% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 74/208	medium · avg 53%	▇▇▇▇ 10%	sharp · at its level · small run · several runs
MMLU-Pro +2	52.3% ✎	███···██···█ beats 5/5	medium · avg 45%	▇▇ 5.1%	at its level · several runs
GPQA (subset unspecified) lab +1	55.9% ✎	▂▁▂▂▄▃▆▃▅█▄▃ beats 58/104	medium · avg 51%	▇▇ 4.6%	sharp · at its level · small run · several runs
MMLU-Redux lab +1	83.7% ✎	▁▁▃▁▂▃▄▅▇▅█▄ beats 15/63	easy · avg 78%	▇ 2.9%	at its level · small run · several runs
INCLUDE lab +1	61.8% ✎	▂····▂▂▄▄▄█▄ beats 13/36	medium · avg 55%	▇ 2.6%	at its level · small run · several runs
MMLU	72.3% ✎	▄█···█·····▄ beats 5/5	easy · avg 74%	▇ 2.1%	at its level
SuperGPQA lab +1	42.7% ✎	▂▁▂▂▅▆▃▇▇▇█▃ beats 13/46	hard · avg 35%	▇ 1.8%	at its level · small run · several runs
IFBench (aggregation unspecified)	30.3% ✎	█··███··█··█ beats 5/5	hard · avg 33%	▇ 1.8%	at its level

22 more tests

test	result	vs the others in its run	difficulty	counts	why
MMLU-ProX lab +1	61.0% ✎	▁·▁▁▁▃▁▃▂█▅▄ beats 4/35	medium · avg 55%	▇ 1.7%	at its level · small run · several runs
Creative Writing v3 lab +1	61.1% ✎	▃····▂▆··▆█▆ beats 0/14	easy · avg 66%	▇ 1.7%	at its level · small run · several runs
WritingBench lab +1	73.5% ✎	▃·▁▂▁··▅█▂▂▁ beats 0/20	easy · avg 71%	▇ 1.6%	at its level · small run · several runs
LiveBench 2024-11-25 lab +1	63.6% ✎	▄··▂·▂·▂█▆▂▂ beats 1/13	medium · avg 62%	▇ 1.6%	at its level · small run · several runs
BRIDGE clinical text, zero-shot average (87 tasks)	38.5% ✎	▁▁▁▁▁▂▂▃▄▆█▃ beats 63/97	hard · avg 32%	▇ 1.5%	blunt · at its level
Multi-IF (aggregation unspecified) lab +1	66.3% ✎	▁▁▁▁▁▁▃▄█▄▁▂ beats 9/41	medium · avg 65%	▇ 1.5%	blunt · at its level · small run · several runs
MMMLU seven languages (Liquid)	60.7% ✎	▄·█····█···▄ beats 5/5	medium · avg 60%	▇ 1.3%	at its level
Open Persian LLM Leaderboard v2 average (21 tasks)	50.5% ✎	▂▄▂▃█▄▃▆▃▅▂▂ beats 33/81	medium · avg 58%	▇ 1.2%	blunt · at its level
Multi-IF three-turn average	75.5% ✎	▄··█▄··▄···▄ beats 5/5	easy · avg 71%	▇ 1.2%	at its level
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	21.5% ✎	▁▁▂▁▃▄█▆▄▅▂▃ beats 5/39	very hard · avg 22%	▇ 1.2%	blunt
OJBench lab	16.1% ✎	▁···▁▃·▅▄▅██ beats 4/26	very hard · avg 13%	▇ 1.0%	small run
Global-MMLU	73.3% ✎	▃▅▃·▃···▅·▅█ beats 4/11	medium · avg 70%	▇ 0.9%	at its level · small run
MGSM six languages, loose (Liquid)	81.8% ✎	▄·▄·█···▄··▄ beats 4/5	easy · avg 82%	▇ 0.7%	typical
MixEval-Hard	38.0% ✎	█····█·····█ beats 2/2	medium · avg 40%	▇ 0.7%	at its level · small run
Open PT-LLM Leaderboard average (9 Portuguese tasks)	76.2% ✎	▃▁▁▁▁▁▁▁▂█▅▁ beats 231/276	medium · avg 73%	▇ 0.6%	at its level · noisy
GSM-Plus +1	56.2% ✎	█·····███·██ beats 1/5	easy · avg 83%	▇ 0.6%	blunt · small run · several runs
ZebraLogic lab	35.2% ✎	█·▁▁▁▇·▄▄▁▄▁ beats 4/24	medium · avg 49%	▇ 0.5%	at its level · noisy · small run
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy	70.8% ✎	▃·▁▁▂▁▂▅█▇▄▁ beats 53/100	medium · avg 68%	▇ 0.5%	at its level · noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy	82.4% ✎	▁▁▁▁▁▁▁▂▄█▄▃ beats 78/146	easy · avg 83%	▇ 0.3%	noisy
Vectara hallucination leaderboard, 2025 dataset (HHEM-2.3, factual consistency rate)	97.3% ✎	▁▁▂▃▄▄█▆▃▂▂▁ beats 111/140	very easy · avg 94%	▇ 0.2%	blunt · maxed out · noisy
Chess Puzzles (Epoch)	1.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 2/135	very hard · avg 5.9%	▇ 0.1%	sharp · too hard for it · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	94.3% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 79/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

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 Qwen3 4B · needs ~2.5 GB · from huggingface.co/Qwen (official)
chat in the terminal
llama cli -hf Qwen/Qwen3-4B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf Qwen/Qwen3-4B-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 Qwen3 4B · needs ~2.5 GB · from huggingface.co/Qwen (official)
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/Qwen/Qwen3-4B-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 Qwen3 4B · needs ~2.91 GB · from ollama.com
chat in the terminal
ollama run qwen3:4b-q4_K_M

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 Qwen3 4B · needs ~2.1 GB, Mac only · from huggingface.co/Qwen (official)
chat in the terminal
mlx_lm.chat --model Qwen/Qwen3-4B-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model Qwen/Qwen3-4B-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3-4B-Thinking-2507 m… 31%, Qwen/Qwen3-4B-Instruct-2507 m… 30%, LFM2 Technical Report — Liqui… 15% and 11 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
Qwen/Qwen3-4B-Thinking-2507 model card · lab-reported	23	▇▇▇▇▇▇▇▇▇ 31%	apache-2.0
Qwen/Qwen3-4B-Instruct-2507 model card · lab-reported	25	▇▇▇▇▇▇▇▇▇ 30%	apache-2.0
LFM2 Technical Report — Liquid AI own SmolLM3 evaluation	11	▇▇▇▇ 15%	CC BY 4.0
HuggingFaceTB/SmolLM3-3B model card (Evaluation section)	8	▇▇▇ 10%	Apache-2.0
Epoch AI — internal evaluations	2	▇ 4%	CC BY 4.0
Arena-Hard-Auto v2.0-Preview leaderboard (repository README)	2	▇ 3%	Apache-2.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)	1	▇ 1%	MIT
Open Persian LLM Leaderboard v2 (Part DP AI and AUT NLP Lab)	1	▇ 1%	Apache-2.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 1%	CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇ 1%	MIT
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)	1	▇ 1%	CC BY 4.0
Berkeley Function Calling Leaderboard result archive	1	▇ 1%	Apache-2.0
Open Portuguese LLM Leaderboard (archived)	1	▇ 1%	Apache-2.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	2	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 50.9