whatmodel.sh / Qwen3-VL 4B Thinking

Qwen3-VL 4B Thinking

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

Qwen3-VL 4B Thinking is an open-weights language model from Alibaba, released October 2025. Its published size is 4.44B parameters and a 256k-token context window. Licence: apache-2.0.

Catalogue identity notes

Maker repo Qwen/Qwen3-VL-4B-Thinking @1de27d8 (licence tag apache-2.0, created 2025-10-11); the separate Thinking checkpoint, not Qwen3-VL-4B-Instruct. Scored from the Qwen3-VL Technical Report (arXiv 2511.21631v2, Tables 3/4 and 7-10: column 'Qwen3-VL 4B thinking'), whose instruct and thinking columns are the two published checkpoints. Released 2025-10: month the maker's Hugging Face repo was created.

Model information sources

Release date ↗ · Weights licence ↗

4,437,815,808 parameters in the Hugging Face safetensors metadata of Qwen/Qwen3-VL-4B-Thinking @1de27d8 (vision encoder included); dense, so active = total.

56.1±0.7#244 of 719 LLMsfrom 27 benchmark results across 3 sources

$ how it compares

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

#modelscorewhereRAMreleased
293Qwen Plus (2025-01-25) Alibaba56±1api $0.60/M2025-01
294Grok 2 (1212) SpaceXAI (xAI)56±1comparison only2024-12
295Qwen3-VL 4B Thinking Alibaba56±1local~4.1GB2025-10
296EXAONE Deep 32B LG AI Research56±2local~22GB2025-03
297DeepSeek R1 Distill Qwen 32B DeepSeek56±1api $1.59/M +local~21GB2025-01
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▁▁▃▆█▆▅▅▅▄▃▁ #179 of 300 * · 2.0% of its score · 1 test

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Mellum2 12B-A2.5B Instruct here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
BFCL v3 lab67.3% ✎█▄··▄█·█▄█▄█ beats 8/13medium · avg 59%▇ 2.0%blunt · at its level
coding▁▁▁▃▆██▇▆▄▃▁ #249 of 458 * · 4.3% of its score · 1 test

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails QwQ-32B-Preview, InternLM2.5 20B Chat, Mistral Small Instruct 22B (2409) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench v6 (window unspecified) lab51.3% ✎▃▅█▅▅▃▅···▃█ beats 9/16medium · avg 35%▇▇ 4.3%at its level
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #201 of 393 * · 2.8% of its score · 1 test

Beats Granite 4.2 8B, Qwen3-VL 30B-A3B Thinking, Mistral Small 4 (119B-A6.5B) 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
IFEval (aggregation unspecified) lab82.6% ✎▃··▃···██▅▃▃ beats 4/11easy · avg 77%▇ 2.8%typical
math▁▂▃▅▆▅▄█▄▁▁▁ #154 of 398 * · 10% of its score · 2 tests

Beats MiMo-V2.5 (310B-A15B), MiniMax-M2.7, MiniMax-M2.1 here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
MathVista (testmini) lab79.5% ✎▃··▃▅·▃█▅▅▅█ beats 10/16medium · avg 69%▇▇ 5.3%at its level
AIME 2025 lab74.5% ✎▃·▃▃█··▃▅▃·█ beats 9/12medium · avg 45%▇▇ 4.6%sharp · at its level
vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #102 of 147 · 51% of its score · 15 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-4o (2024-08-06) here, though they rank higher overall. Trails Llama 4 Scout (109B-A17B), Mistral Small 3.2 24B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
CharXiv (reasoning questions) lab50.3% ✎▄▂▄▆▄▄▆··▄█▂ beats 8/21medium · avg 39%▇▇▇▇ 9.2%at its level
MMMU (val) lab70.8% ✎▂···▆▂▂▂▄▄▄█ beats 10/16medium · avg 61%▇▇▇ 7.7%at its level
CharXiv descriptive questions lab83.9% ✎▄··▂▂▂·▂▄▄▂█ beats 6/14easy · avg 74%▇▇ 6.2%typical
MathVista (testmini) lab79.5% ✎▃··▃▅·▃█▅▅▅█ beats 10/16medium · avg 69%▇▇ 5.3%at its level
MMStar lab73.2% ✎▂··▄▂▂·▄█▃▄▆ beats 11/16medium · avg 65%▇▇ 5.0%at its level
MMMU Pro lab57.0% ✎▃·█·▅▃▃▃▅▅▃█ beats 10/16medium · avg 46%▇▇ 4.4%at its level
ChartArena chart parsing (English, mAP_high average)34.7% ✎▁·▁·▃▃▁▇▃█·▄ beats 9/22hard · avg 27%▇ 2.5%blunt · at its level
HallusionBench lab64.1% ✎▃·▃·▅▃▅▅▃▅█▅ beats 13/16medium · avg 55%▇ 2.0%blunt · at its level
7 more tests
testresultvs the others in its rundifficultycountswhy
BEAR-Bench (bilingual RU/EN enterprise and academic document images, overall accuracy)34.5% ✎▃▃·▅·▃·█·█▅▅ beats 4/14hard · avg 24%▇ 1.7%at its level
ChartQA lab88.8% ✎▂··▂·▂▃·▅▂▆█ beats 11/17easy · avg 85%▇ 1.5%blunt
RealWorldQA lab73.2% ✎▅▃·▃▅▅·█·▃▅█ beats 8/16easy · avg 70%▇ 1.4%blunt · at its level
DocVQA lab94.2% ✎▃··▅▃▅▃█▃█▅▅ beats 4/17very easy · avg 94%▇ 1.3%blunt · maxed out
ScreenSpot-Pro lab49.2% ✎▂····▂▄··▂█▆ beats 3/11medium · avg 46%▇ 1.2%blunt · at its level
CC-OCR lab73.8% ✎▃··▃▅·▅▅·█▅▃ beats 2/13easy · avg 74%▇ 1.0%blunt
OCRBench lab80.8% ✎▅█·▃▃█▃▅▅▃·▃ beats 2/16easy · avg 84%▇ 0.5%blunt
other tests8 not in a task · 35% of its score
testresultvs the others in its rundifficultycountswhy
GPQA (subset unspecified) lab64.1% ✎▂····▂·▂▂█▆█ beats 7/14hard · avg 49%▇▇▇ 8.7%sharp · at its level
MMLU-Pro lab73.6% ✎▄·▄·▄▄▄▄█▄██ beats 6/12medium · avg 63%▇▇▇ 6.8%at its level
MMLU-Redux lab86.0% ✎▃··▃·▃·█▃▅▃▅ beats 5/11easy · avg 79%▇▇ 5.5%typical
SuperGPQA lab46.8% ✎▄··▄·▄▄▄█▄██ beats 5/11medium · avg 37%▇▇ 4.3%at its level
AndroidWorld lab52.0% ✎▃··▃▅▃▃▃▃▃·█ beats 5/11medium · avg 43%▇ 3.1%blunt · at its level
LVBench lab53.5% ✎▅··▃·▅▃▃·█·▃ beats 2/10hard · avg 51%▇ 3.0%at its level
Video-MME (without subtitles) lab68.9% ✎▃··█··▅█▃▃▅▃ beats 4/13medium · avg 66%▇ 2.7%blunt · at its level
MVBench lab69.3% ✎▂▂▂·▂█··▆▄·▂ beats 7/13medium · avg 67%▇ 1.1%blunt · at its level

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

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/Qwen/Qwen3-VL-4B-Thinking-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-VL 4B Thinking · needs ~3.3 GB · from ollama.com
    chat in the terminal
    ollama run qwen3-vl:4b-thinking

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-VL 4B Thinking · needs ~3.1 GB, Mac only · from huggingface.co/lmstudio-community
    chat in the terminal
    mlx_lm.chat --model lmstudio-community/Qwen3-VL-4B-Thinking-MLX-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model lmstudio-community/Qwen3-VL-4B-Thinking-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen (Alibaba) vendor-reporte… 96%, ChartArena: Benchmarking Char… 2%, BEAR-Bench: A Bilingual Enter… 2%

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 (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card) · lab-reported25▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 96%Factual vendor-reported score (not licensed content)
ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)1▇ 2%CC BY 4.0
BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)1▇ 2%CC BY 4.0
prior · pull toward the middle▇ 0%
weighted average= 56.1

$ 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