Qwen /

Qwen2.5-VL-7B-Instruct

A 7B model that reads images, charts, documents and video, and answers in text.

1.8k likes
5.1M downloads a month
8.3B parameters
Some experience
Vision & language

Open on Hugging Face (opens in a new tab)

Qwen2.5-VL-7B-Instruct is a multimodal model: you give it pictures, several images, or a video along with a question, and it writes an answer. The card says it can recognise common objects and read text, charts, icons and layouts in images. It can also point to objects by giving bounding boxes or points, and it can produce structured output (such as JSON) for things like invoices, forms and tables.

What sets it apart, according to the card, is its range: it can handle videos over an hour long and pinpoint the relevant segment, and it is described as a visual agent that can direct tools, including computer use and phone use. The honest limits: the card's own agent scores show room to grow (for example AndroidWorld_SR 25.5 and ScreenSpot Pro 29.0). The default context is 32,768 tokens, and the YaRN method for going longer is not recommended because it hurts localization tasks. The card also says videos currently only work from local files in the usage tips, while its video table shows URL support depends on which backend you use.

What you need to run it

8.3B parameters. The weights take about 17 GB at full 16-bit precision (fits a 24 GB Mac or a GPU with 24 GB), or about 5 GB as a 4-bit download, which runs on a 8 GB Mac or a GPU with 8 GB. Long inputs and images need more memory on top.

What it does well

Reads documents and charts
On the card's benchmarks it scores 95.7 on DocVQA, 82.6 on InfoVQA, 87.3 on ChartQA and 864 on OCRBench, which are the best numbers in the compared columns.
Long video understanding
The card says it can comprehend videos of over 1 hour and pinpoint the relevant segments. It uses dynamic frame-rate sampling so it can learn time order and speed.
Locating things in images
It can mark objects with bounding boxes or points and give stable JSON output for coordinates and attributes.
Structured output
It can turn scans of invoices, forms and tables into structured data, which the card suggests for finance and commerce.
Adjustable image cost
You can set min_pixels and max_pixels (the card gives a token range of 256-1280 as an example) to trade quality against speed and memory.

What you could build

Ask questions about a screenshot, chart or photo of a worksheet and get a text answer.
Pull the fields out of an invoice or form scan into JSON for a small bookkeeping tool.
Build a tool that finds and boxes objects in an image and returns their coordinates.
Describe or search a long recorded lecture or video by asking about specific moments.
Compare two or more images and list their similarities, as in the card's multi-image example.

Get started

Install the latest transformers from source plus the qwen-vl-utils helper, then follow the card's Quickstart, which loads the model with Qwen2_5_VLForConditionalGeneration and asks it to describe an image. On non-Linux systems the card says decord may not install and suggests the plain qwen-vl-utils package instead.

pip install git+https://github.com/huggingface/transformers accelerate

pip install qwen-vl-utils[decord]==0.0.8

pip install qwen-vl-utils

Commands are copied from the model card. Read it on Hugging Face for the full setup.