whatmodel.sh / Qwen3 4B Thinking 2507

Qwen3 4B Thinking 2507

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Qwen3 4B Thinking 2507 is an open-weights language model from Alibaba, released August 2025. Its published size is 4.02B parameters and a 262k-token context window. Licence: apache-2.0.

56.6±0.8#240 of 719 LLMsfrom 33 benchmark results across 8 sources

$ how it compares

The LLMs ranked either side of Qwen3 4B Thinking 2507, out of 828 with a score, then more from Alibaba.

#modelscorewhereRAMreleased
285Claude 3.5 Sonnet (2024-06-20) Anthropic57±1comparison only2024-06
286Qwen2.5 Coder 32B Instruct Alibaba57±1api $0.74/M +local~21GB2024-11
287Qwen3 4B Thinking 2507 Alibaba57±1local~3.9GB2025-08
288Gemini 1.5 Pro Experimental (0827) Google56±1comparison only2024-08
289Gemini 1.5 Pro (002) Google56±1comparison only2024-09
more from Alibaba
18Qwen3.8 Max (2.4T-A95B) Alibaba78±1api $3.00/M +local~380GB2026-08
34Qwen3.8 Max (0902) Alibaba76±1api $3.00/M2026-09
37Qwen3.8-Flash-Next Alibaba75±1local~59GB2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #166 of 300 · 19% of its score · 6 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Jamba 1.5 Large here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
τ²-bench Telecom lab27.2% ✎▁·▅█▆▅▂▃▃▂▃▁ beats 20/89very hard · avg 14%▇▇▇ 8.4%sharp · at its level · small run
τ-bench Retail lab66.1% ✎▁····▃·▂▇█▂▆ beats 15/28medium · avg 42%▇ 2.9%at its level · small run
τ-bench Airline lab48.0% ✎▁▁▁·▅▄▁█▇▁▁▇ beats 17/28hard · avg 28%▇ 2.5%at its level · small run
τ²-bench Retail lab53.5% ✎▁····▂▁▃▆█▃▁ beats 6/44hard · avg 35%▇ 2.4%at its level · small run
τ²-bench Airline lab58.0% ✎▁▁·▂▁▃▂▁█▅▂▃ beats 25/42hard · avg 31%▇ 1.6%at its level · small run
BFCL v3 lab71.2% ✎▁▁▁▁▁▃▅█▆▅▄▁ beats 155/170medium · avg 61%▇ 0.9%blunt · at its level · small run
coding▁▁▁▃▆██▇▆▄▃▁ #223 of 458 * · 2.8% of its score · 1 test

Beats Qwen3.5 4B, Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B) here, though they rank higher overall. Trails QwQ-32B-Preview, Ministral 3 3B Reasoning 2512, GPT-4 (0613) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench v6 (Feb–May 2025) lab55.2% ✎▁▁▃▁█▅▄▇▅▅▃█ beats 8/38medium · avg 43%▇ 2.8%at its level · small run
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #184 of 393 * · 6.6% of its score · 2 tests

Beats Seed-OSS 36B Instruct, Nemotron-Cascade-14B-Thinking, Granite 4.2 8B here, though they rank higher overall. Trails LFM2.5 1.2B Instruct, LFM2.5 1.2B Thinking, Gemma 3 12B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena-Hard v2 aggregate lab34.9% ✎▁▁··▄▃▁▃█▁▁▇ beats 2/22hard · avg 20%▇▇ 5.1%sharp · at its level · small run
IFEval (aggregation unspecified) lab87.4% ✎▁▁▃▁▂▄▄▆█▃▂▁ beats 88/141easy · avg 80%▇ 1.5%small run
math▁▂▃▅▆▅▄█▄▁▁▁ #181 of 398 · 27% of its score · 6 tests

Beats MiniMax-M2.1, GPT-4.1, DeepSeek V3 0324 here, though they rank higher overall. Trails Phi-4-mini-reasoning 3.8B, OpenThinker3 7B, Olmo 3 7B Think here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
HMMT 2025 (edition unspecified) lab55.5% ✎▃█▇▅▇▅▁·▁··▁ beats 13/32hard · avg 34%▇▇▇ 6.7%sharp · at its level · small run
MathArena ArXivMath January 2026 (23 problems)23.9% ✎▃··▄▂▄█▆▆▄▄█ beats 0/23hard · avg 22%▇▇ 5.4%at its level
AIME 2025 lab81.3% ✎▁·▁▁▁▂▅▆█▇▅▁ beats 87/133medium · avg 59%▇▇ 5.3%sharp · small run
MathArena ArXivMath February 2026 (32 problems)18.0% ✎▁▃▁▁█▇▄▃▁▃·▃ beats 1/24very hard · avg 9.1%▇▇ 4.5%typical
MathArena ArXivMath December 2025 (17 problems)32.4% ✎▃·▃▃█▁·▄·▁▁▁ beats 2/18hard · avg 19%▇ 3.1%at its level
PolyMATH lab46.2% ✎▁···▃▁·▇█▄▄▄ beats 8/23hard · avg 34%▇ 2.3%at its level · small run
other tests17 not in a task · 44% of its score
testresultvs the others in its rundifficultycountswhy
MMLU-Pro lab +174.0% ✎▁▁·▁▁▁▃▂▄▅█▁ beats 98/188medium · avg 65%▇▇ 6.2%at its level · several runs
GPQA (subset unspecified) lab65.8% ✎▂▁▂▂▄▃▆▃▅█▄▃ beats 68/104medium · avg 51%▇▇ 5.9%sharp · at its level · small run
Medmarks-V weighted mean win rate (61 models, 71 configurations)48.9% ✎▂▃▅▂▆▃█▅█▅▃▆ beats 15/34medium · avg 42%▇▇ 4.8%blunt · at its level
BRIDGE clinical text, zero-shot average (87 tasks)40.0% ✎▁▁▁▁▁▂▂▃▄▆█▃ beats 76/97hard · avg 32%▇ 3.5%blunt · at its level
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items24.8% ✎▁▁▂▁▃▄█▆▄▅▂▃ beats 9/39very hard · avg 22%▇ 3.2%blunt
INCLUDE lab64.4% ✎▂····▂▂▄▄▄█▄ beats 15/36medium · avg 59%▇ 2.7%at its level · small run
OJBench lab17.9% ✎▁···▁▃·▅▄▅██ beats 5/26very hard · avg 13%▇ 2.7%small run
MMLU-Redux lab86.1% ✎▁▁▃▁▂▃▄▅▇▅█▄ beats 23/63easy · avg 81%▇ 2.5%small run
9 more tests
testresultvs the others in its rundifficultycountswhy
SuperGPQA lab47.8% ✎▂▁▂▂▅▆▃▇▇▇█▃ beats 18/46medium · avg 39%▇ 2.0%at its level · small run
Creative Writing v3 lab75.6% ✎▂····▃▆··▆█▆ beats 1/14easy · avg 66%▇ 1.9%small run
MMLU-ProX lab64.2% ✎▁·▁▁▁▃▁▃▂█▅▄ beats 6/35medium · avg 59%▇ 1.7%at its level · small run
WritingBench lab83.3% ✎▁·▁▂▁··▅█▂▂▁ beats 5/20easy · avg 71%▇ 1.7%small run
LiveBench 2024-11-25 lab71.8% ✎▄··▂·▃·▂█▆▂▂ beats 3/13medium · avg 62%▇ 1.6%at its level · small run
Multi-IF (aggregation unspecified) lab77.3% ✎▁▁▁▁▁▁▃▄█▄▁▂ beats 24/41medium · avg 65%▇ 1.4%blunt · at its level · small run
Open PT-LLM Leaderboard average (9 Portuguese tasks)59.1% ✎▃▁▁▁▁▁▁▁▃█▅▁ beats 87/276medium · avg 73%▇ 1.2%at its level · noisy
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy74.0% ✎▃·▁▁▂▁▂▅█▇▄▁ beats 58/100medium · avg 68%▇ 0.9%at its level · noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy89.1% ✎▁▁▁▁▁▁▁▂▄█▄▃ beats 113/146easy · avg 83%▇ 0.5%noisy

$ where to run it

llama.cpp · recommended

version

  1. 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 ↗

  2. Run Qwen3 4B Thinking 2507 · needs ~2.5 GB · from huggingface.co/lmstudio-community
    chat in the terminal
    llama cli -hf lmstudio-community/Qwen3-4B-Thinking-2507-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf lmstudio-community/Qwen3-4B-Thinking-2507-GGUF:Q4_K_M

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run Qwen3 4B Thinking 2507 · needs ~2.5 GB · from huggingface.co/lmstudio-community

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/lmstudio-community/Qwen3-4B-Thinking-2507-GGUF@Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

  1. Install Ollama

    download it (macOS / Windows) ↗

    Linux
    curl -fsSL https://ollama.com/install.sh | sh
  2. Run Qwen3 4B Thinking 2507 · needs ~2.91 GB · from ollama.com
    chat in the terminal
    ollama run qwen3:4b

MLX (Mac)

version

  1. Install MLX (Mac) · Apple silicon Macs only, text only.
    with uv
    uv tool install mlx-lm
    or
    pip install mlx-lm
  2. Run Qwen3 4B Thinking 2507 · needs ~2.3 GB, Mac only · from huggingface.co/lmstudio-community
    chat in the terminal
    mlx_lm.chat --model lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3-4B-Thinking-2507 m… 69%, MathArena ArXivMath / BrokenA… 13%, Medmarks: A Comprehensive Ope… 5% 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

sourceresultsshare of the scorelicence
Qwen/Qwen3-4B-Thinking-2507 model card · lab-reported23▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 69%apache-2.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets3▇▇▇▇ 13%CC BY-SA 4.0
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models1▇ 5%CC BY 4.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)1▇ 3%MIT
Qwen/Qwen3.5-0.8B model card · lab-reported1▇ 3%apache-2.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)1▇ 3%CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard2▇ 1%MIT
Open Portuguese LLM Leaderboard (archived)1▇ 1%Apache-2.0
prior · pull toward the middle▇ 0%
weighted average= 56.6
$ quantization evidence · 12 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.

precisiontestfull → quantizedkept
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)MATH-500 — published row 'Math 500'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.894 → 0.906101.3%source ↗
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)IFEval — published row 'IfEval (0-shot)'; lm-evaluation-harness task ifeval via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.902 → 0.902100.0%source ↗
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)GSM8K-Platinum — published row 'GSM8k Platinum (0-shot)'; lm-evaluation-harness task gsm8k_platinum_cot_llama via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.953 → 0.94999.6%source ↗
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)AIME 2025 — published row 'AIME 2025'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.738 → 0.72197.7%source ↗
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)MMLU-Pro — published row 'MMLU-Pro (0-shot)'; lm-evaluation-harness task mmlu_pro_chat via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.736 → 0.71396.9%source ↗
INT4 W4A16 (weight-only INT4, symmetric, group 128; GPTQ, llm-compressor)GPQA Diamond — published row 'GPQA diamond'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.689 → 0.63391.8%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)GSM8K-Platinum — published row 'GSM8k Platinum (0-shot)'; lm-evaluation-harness task gsm8k_platinum_cot_llama via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.953 → 0.958100.5%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)MATH-500 — published row 'Math 500'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.894 → 0.89399.8%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)IFEval — published row 'IfEval (0-shot)'; lm-evaluation-harness task ifeval via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.902 → 0.89899.6%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)MMLU-Pro — published row 'MMLU-Pro (0-shot)'; lm-evaluation-harness task mmlu_pro_chat via vLLM OpenAI-compatible server (backend local-chat-completions, chat template, --num_fewshot 0, gen_kwargs do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000)0.736 → 0.72798.8%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)AIME 2025 — published row 'AIME 2025'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.738 → 0.71797.2%source ↗
INT8 W8A8 (weights INT8 per-channel, activations INT8 per-token dynamic; SmoothQuant+GPTQ, llm-compressor)GPQA Diamond — published row 'GPQA diamond'; lighteval (litellm endpoint to vLLM server), temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 320000.689 → 0.66796.7%source ↗

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit 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 · sources and licences · json