whatmodel.sh / Qwen3-VL 32B Thinking

Qwen3-VL 32B Thinking

+ contribute

Qwen3-VL 32B Thinking is an open-weights language model from Alibaba, released October 2025. Its published size is 33.4B parameters and a 256k-token context window. Licence: apache-2.0.

Catalogue identity notes

Comparison-only, proposed so the HyperCLOVA X 32B Think report's Terminal Bench Hard run (Table 5) has two catalogue models (with exaone-4-0-32b). The report column 'Qwen3-VL 32B-Thinking' (cited Yang et al., 2025). Maker repo Qwen/Qwen3-VL-32B-Thinking @7edd10f: licence tag apache-2.0, created 2025-10-19, 33,357,390,064 parameters in safetensors metadata (dense). Separate Thinking checkpoint (not Qwen3-VL-32B-Instruct). Card: 'Native 256K context, expandable to 1M'.

61.9±0.7#180 of 719 LLMsfrom 35 benchmark results across 3 sources

$ how it compares

The LLMs ranked either side of Qwen3-VL 32B Thinking, out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
209	gpt-oss-120b OpenAI	62±1	api $0.26/M +local	~67GB	2025-08
210	GPT-5 nano OpenAI	62±1	api $0.14/M		2025-08
211	Qwen3-VL 32B Thinking Alibaba	62±1	api $0.28/M +local	~22GB	2025-10
212	Gemini 2.0 Flash-Lite (001) Google	62±2	comparison only		2025-02
213	Mistral Large 3 (675B-A41B) Mistral	62±1	api $0.75/M +local	~108GB	2025-12
	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▁▁▃▆█▆▅▅▅▄▃▁ #155 of 300 · 20% of its score · 5 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.

test	result	vs the others in its run	difficulty	counts	why
τ²-bench Telecom lab +1	46.9% ✎	▃·█········█ beats 6/6	very hard · avg 10%	▇▇▇▇ 11%	sharp · at its level · several runs
Terminal-Bench Hard	6.9% ✎	▂▅▆▂▃▆▂▂█▅▅▂ beats 9/27	very hard · avg 2.3%	▇ 3.4%	sharp · small run
τ²-bench Retail	57.0% ✎	▁····▂▁▃▆█▃▁ beats 7/44	medium · avg 38%	▇ 2.2%	at its level
τ²-bench Airline	54.4% ✎	▁▁·▂▁▃▂▃█▅▂▃ beats 15/42	hard · avg 29%	▇ 1.8%	at its level
BFCL v3 lab	71.7% ✎	█▄··▄█·█▄█▄█ beats 12/13	medium · avg 59%	▇ 1.7%	blunt · at its level

coding▁▁▁▃▆██▇▆▄▃▁ #174 of 458 * · 3.7% of its score · 1 test

Beats Qwen3.5 9B, Devstral 2 (2512) here, though they rank higher overall. Trails QwQ-32B-Preview, Ministral 3 8B Reasoning 2512, Phi-4-reasoning 14B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench v6 (window unspecified) lab	65.6% ✎	▃▅█▅▅▃▅···▃█ beats 12/16	medium · avg 35%	▇ 3.7%	at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #151 of 393 * · 2.0% of its score · 1 test

Beats INTELLECT-3, MiniMax-M2, MiniCPM5 2B here, though they rank higher overall. Trails LFM2.5 1.2B Thinking, LFM2.5 8B-A1B, LFM2.5 2.6B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
IFEval (aggregation unspecified) lab	87.8% ✎	▃··▃···██▅▃▃ beats 11/11	easy · avg 77%	▇ 2.0%	typical

math▁▂▃▅▆▅▄█▄▁▁▁ #105 of 398 · 9.3% of its score · 4 tests

Beats Claude Sonnet 4.6, MiniMax-M3, Nemotron 3 Ultra 550B-A55B here, though they rank higher overall. Trails Gemini 2.5 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MathVista (testmini) lab	85.9% ✎	▃··▃▅·▃█▅▅▅█ beats 16/16	medium · avg 69%	▇▇ 3.9%	typical
AIME 2025 lab	83.7% ✎	▃·▃▃█··▃▅▃·█ beats 12/12	medium · avg 45%	▇ 3.2%	sharp
OMEGA (math generalisation)	50.8% ✎	▅···▃▃···▃▃█ beats 6/8	hard · avg 24%	▇ 1.6%	at its level
MATH (subset unspecified)	96.7% ✎	▅····▃▃··▃▃█ beats 8/8	easy · avg 74%	▇ 0.5%	noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #83 of 147 · 38% of its score · 13 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-5 mini, GLM-4.6V here, though they rank higher overall. Trails GPT-4.1 mini, Gemini 2.5 Flash, Mistral Large 3 (675B-A41B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
CharXiv (reasoning questions) lab	65.2% ✎	▄▂▄▆▄▄▆··▄█▂ beats 13/21	medium · avg 39%	▇▇▇ 8.0%	at its level
MMMU (val) lab	78.1% ✎	▂···▆▂▂▂▄▄▄█ beats 15/16	medium · avg 61%	▇▇ 6.2%	at its level
MMStar lab	79.4% ✎	▂··▄▂▂·▄█▂▄▆ beats 16/16	medium · avg 65%	▇▇ 4.1%	at its level
CharXiv descriptive questions lab	90.2% ✎	▄··▂▂▂·▂▄▄▂█ beats 12/14	easy · avg 74%	▇▇ 4.1%	typical
MathVista (testmini) lab	85.9% ✎	▃··▃▅·▃█▅▅▅█ beats 16/16	medium · avg 69%	▇▇ 3.9%	typical
MMMU Pro lab	68.1% ✎	▃·█·▅▃▃▃▅▅▃█ beats 15/16	medium · avg 46%	▇▇ 3.9%	at its level
HallusionBench lab	67.4% ✎	▃·▃·▅▃▅▅▃▅█▅ beats 16/16	medium · avg 55%	▇ 1.8%	blunt · at its level
RealWorldQA lab	78.4% ✎	▅▃·▃▅▅·█·▃▅█ beats 14/16	easy · avg 70%	▇ 1.2%	blunt

5 more tests

test	result	vs the others in its run	difficulty	counts	why
ChartQA lab	89.0% ✎	▂··▂·▂▃·▅▂▆█ beats 12/17	easy · avg 85%	▇ 1.2%	blunt
ScreenSpot-Pro lab	57.1% ✎	▂····▂▄··▂█▆ beats 5/11	medium · avg 46%	▇ 1.1%	blunt · at its level
DocVQA lab	96.1% ✎	▃··▅▃▅▃█▃█▅▅ beats 12/17	very easy · avg 94%	▇ 1.0%	blunt · maxed out
CC-OCR lab	79.6% ✎	▃··▃▅·▅▅·█▅▃ beats 8/13	easy · avg 74%	▇ 0.8%	blunt
OCRBench lab	85.5% ✎	▅█·▃▃█▃▅▅▃·▃ beats 6/16	easy · avg 84%	▇ 0.4%	blunt

other tests11 not in a task · 31% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA (subset unspecified) lab	73.1% ✎	▂····▂·▂▂█▆█ beats 11/14	hard · avg 49%	▇▇▇ 7.2%	sharp · at its level
MMLU-Pro lab	82.1% ✎	▄·▄·▄▄▄▄█▄██ beats 12/12	medium · avg 63%	▇▇ 4.9%	typical
SuperGPQA lab	59.0% ✎	▄··▄·▄▄▄█▄██ beats 11/11	medium · avg 37%	▇▇ 3.9%	at its level
MMLU-Redux lab	91.9% ✎	▃··▃·▃·█▃▅▃▅ beats 11/11	easy · avg 79%	▇ 3.7%	typical
LVBench lab	62.6% ✎	▅··▃·▅▃▃·█·▃ beats 8/10	hard · avg 51%	▇ 2.8%	at its level
AndroidWorld lab	63.7% ✎	▃··▃▅▃▃▃▃▃·█ beats 10/11	medium · avg 43%	▇ 2.8%	blunt · at its level
Video-MME (without subtitles) lab	77.3% ✎	▃··█··▅█▃▃▅▃ beats 12/13	medium · avg 66%	▇ 2.3%	blunt · at its level
HellaSwag	62.1% ✎	▁▁▁▁▂▂▃▇█▆▅▃ beats 23/85	medium · avg 64%	▇ 1.3%	at its level · small run

3 more tests

test	result	vs the others in its run	difficulty	counts	why
MVBench lab	73.2% ✎	▂▂▂·▂█··▆▄·▂ beats 11/13	medium · avg 67%	▇ 1.0%	blunt · at its level
MMLU	71.5% ✎	▁▁▁▁▁▂▃▅█▃▆▃ beats 120/183	medium · avg 71%	▇ 0.8%	noisy · small run
PIQA	88.0% ✎	▁▁▁▃▃▄▇█▄▃▁▃ beats 82/83	easy · avg 84%	▇ 0.3%	blunt · small run

$ where to run it

Through an API

api · Alibaba Cloud Model Studio · $0.28/M · $0.16 in / $0.64 out per 1M tokens · 256k context · www.alibabacloud.com ↗ · price sourceUSD per 1M tokens, Alibaba Cloud Model Studio, API model qwen3-vl-32b-thinking ('Thinking mode only', deployment scope 'Global'): input $0.16, output $0.64, the same in both regional tables that list it and equal to the catalogued qwen3-vl-32b-instruct row. Free quota, batch and context-cache discounts excluded. Observed 2026-10-11.

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-VL 32B Thinking · needs ~19.8 GB · from huggingface.co/Qwen (official)
chat in the terminal
llama cli -hf Qwen/Qwen3-VL-32B-Thinking-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf Qwen/Qwen3-VL-32B-Thinking-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-VL 32B Thinking · needs ~19.8 GB · from huggingface.co/Qwen (official)
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-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-VL 32B Thinking · needs ~21 GB · from ollama.com
chat in the terminal
ollama run qwen3-vl:32b-thinking

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-VL 32B Thinking · needs ~19.6 GB, Mac only · from huggingface.co/lmstudio-community
chat in the terminal
mlx_lm.chat --model lmstudio-community/Qwen3-VL-32B-Thinking-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model lmstudio-community/Qwen3-VL-32B-Thinking-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen (Alibaba) vendor-reporte… 85%, HyperCLOVA X 32B Think (NAVER… 13%, allenai/Olmo-3.1-32B-Think mo… 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

source	results	share of the score	licence
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-reported	26	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 85%	Factual vendor-reported score (not licensed content)
HyperCLOVA X 32B Think (NAVER Cloud, arXiv:2601.03286v1)	7	▇▇▇▇ 13%	CC-BY-4.0
allenai/Olmo-3.1-32B-Think model card	2	▇ 2%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 61.9

$ 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