Qwen2.5-VL-32B-Instruct
Alibaba · Qwen/Qwen2.5-VL-32B-Instruct
Weights license: Apache 2.0 · code and components may use other terms
Quality
Quality · LMArena · vision
LMArena 2026-10-19 · human-preference rating · higher is better · this model outlined
- Color = publisher
Source: LMArena 2026-10-19 · collected 2026-10-25 · dataset · CC BY 4.0 · models without a rating are not drawn (no rating ≠ 0)
1,119 ± 15 · LMArena · vision · Confidence interval 1,103–1,134 · 1,490 votes
Reported by the publisher
- MMMU
- 70Qwen2.5-VL-32B (🤗🤖) · Vision
- MMMU Pro
- 49.5Qwen2.5-VL-32B (🤗🤖) · Vision
- MMStar
- 69.5Qwen2.5-VL-32B (🤗🤖) · Vision
- MathVista
- 74.7Qwen2.5-VL-32B (🤗🤖) · Vision
- MathVision
- 40.0Qwen2.5-VL-32B (🤗🤖) · Vision
- CC-OCR
- 77.1Qwen2.5-VL-32B (🤗🤖) · Vision
- DocVQA
- 94.8Qwen2.5-VL-32B (🤗🤖) · Vision
- InfoVQA
- 83.4Qwen2.5-VL-32B (🤗🤖) · Vision
Every reported result
- LVBench
- 49.00Qwen2.5-VL-32B (🤗🤖) · Vision
- CharadesSTA
- 54.2Qwen2.5-VL-32B (🤗🤖) · Vision
- MMBench-Video
- 1.93Qwen2.5-VL-32B (🤗🤖) · Vision
- AITZ
- 83.1Qwen2.5-VL-32B (🤗🤖) · Vision
- ScreenSpot
- 88.5Qwen2.5-VL-32B (🤗🤖) · Vision
- ScreenSpot Pro
- 39.4Qwen2.5-VL-32B (🤗🤖) · Vision
- AndroidWorld
- 22.0Qwen2.5-VL-32B (🤗🤖) · Vision
- OSWorld
- 5.92Qwen2.5-VL-32B (🤗🤖) · Vision
- MMLU
- 78.4Qwen2.5-VL-32B · Text
- MMLU-PRO
- 68.8Qwen2.5-VL-32B · Text
- MATH
- 82.2Qwen2.5-VL-32B · Text
- GPQA-diamond
- 46.0Qwen2.5-VL-32B · Text
- MBPP
- 84.0Qwen2.5-VL-32B · Text
- Human Eval
- 91.5Qwen2.5-VL-32B · Text
Values as printed on the card, with the publisher's own settings (card section: Evaluation › Vision; Evaluation › Text). Not comparable across publishers and not used in any ranking. Model card
Quality × memory
No comparable quality × memory pair
This entry has no score on a benchmark that is drawn against planning memory.
Alibaba released Qwen2.5-VL-32B-Instruct on 2025-03-25: a vision-language model that answers about images, with 33.5B parameters on the Hugging Face count and the Apache 2.0 license. Languages the publisher declares: en.
License: commercial use allowed. At Q4 the weights load about 25.1 GB before each machine's reserve; image resolution, video frames and context add memory, and the vision encoder needs support in the runtime you pick.
Runs on
Fast memory holds it 33
The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.
Show all 33
Within memory, outside the fast part 51
Within the machine's memory, outside its fast part: on a graphics card the layers that do not fit run on the CPU from PC RAM (the listed RAM, or 32 GB assumed); on a Mac, above the macOS default GPU share. Not a speed figure.
Show all 51
One memory tier short 9
The same machines with the next memory size up would hold it in fast memory.
Show all 9
Hardware profiles grouped by the Q4 planning estimate on each machine (dedicated GPUs reserve 2 GB; unified and system memory 25%, 2 to 8 GB): the fast memory holds it, it is within memory but outside the fast part (layers offloaded to PC RAM, or above the macOS GPU share), or the machine is one memory tier short. Not a measured fit or speed.
Measured speed
No published measurement for this checkpoint
Tokens per second are shown only when a source publishes them with hardware, build and setup. They are never estimated.
How to run
Runtime compatibility and tokens per second have not been measured for this entry. Published metadata and a memory estimate do not prove a successful local run.
Library in source: transformers
Documented ways to run it
Copied as printed in the linked source (read 2026-10-18); not run by this site.
# It's highly recommanded to use `[decord]` feature for faster video loading.
pip install qwen-vl-utils[decord]==0.0.8from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
# default: Load the model on the available device(s)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen2.5-VL-32B-Instruct", torch_dtype="auto", device_map="auto"
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
# "Qwen/Qwen2.5-VL-32B-Instruct",
# torch_dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# device_map="auto",
# )
# default processer
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-32B-Instruct")
# The default range for the number of visual tokens per image in the model is 4-16384.
# You can set min_pixels and max_pixels according to your needs, such as a token range of 256-1280, to balance performance and cost.
# min_pixels = 256*28*28
# max_pixels = 1280*28*28
# processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-32B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
},
{"type": "text", "text": "Describe this image."},
],
}
]
# Preparation for inference
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)Technical details
What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.
- Checkpoint / source
- Qwen/Qwen2.5-VL-32B-Instruct
- Task
- image-text-to-text
- Architecture
- Qwen2_5_VLForConditionalGeneration
- Library in source
- transformers
- Weight license
- Apache 2.0Qwen/Qwen2.5-VL-32B-Instruct model card (Hugging Face) · read 2026-10-18source
- Commercial use
- Allowed Apache License 2.0 text (SPDX) · read 2026-10-18license text
- Languages
- enModel card metadata · read 2026-10-25source
- Released
- Qwen2.5-VL README news (GitHub) · read 2026-10-18source
Every stored field
- Revision checked
- 7cfb30d71a1f4f49a57592323337a4a4727301da
- Architecture
- Qwen2_5_VLForConditionalGeneration
- Library
- transformers
- Hub task tag
- image-text-to-text
- Formats
- transformers, safetensors
- Access
- Open download
- Publisher kind
- Official
- Hub parameter count
- 33,452,718,336safetensors.total
- Repository created
- 2025-03-21creation is not the model's release
- Repository updated
- 2025-04-14a commit is not a release date
- Hub downloads (30 days) · likes
- 836,945 · 507this repository only · read 2026-10-25
- Catalog state
- Hub config or weights found; no inventory benchmark link
Sources
Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.
- Publisher model card
- License: Qwen/Qwen2.5-VL-32B-Instruct model card (Hugging Face)
- Release date: Qwen2.5-VL README news (GitHub)
Curated facts checked: 2026-10-18. The license and release date cite the publisher's sources; parameters, files and task come from the Hugging Face record, and memory is a labelled estimate from its parameter count. A value no source states stays empty; a repository commit is not a model release date.


























































