Qwen2.5-VL
Vision-language model for images, documents and long videos
Qwen2.5-VL is a multimodal language model series released in February 2025 by the Qwen team. It processes text, images and videos up to an hour long, supporting object localisation and structured data extraction. The largest variant has 72 billion parameters and aims to match leading closed models in document understanding.
- Processes videos over one hour with event localisation
- Extracts structured data from invoices and forms
- Localises objects using bounding boxes or points
- Available in 3B, 7B and 72B parameter sizes
- Acts as a visual agent for computer and phone use
What it is
Qwen2.5-VL is a family of vision-language models designed to understand and interact with visual content. Released in February 2025, it comes in three sizes: 3, 7 and 72 billion parameters. The system handles static images, complex documents and videos lasting over an hour. It supports tasks such as reading charts, extracting data from forms and locating objects within a scene. The largest model aims to perform at a level comparable to other top-tier systems in visual reasoning and document analysis.
How it came about
This series follows Qwen2-VL, which had been available for five months prior. Developers built many downstream projects on that earlier version, providing feedback that shaped the new release. The team focused on making the models more useful for real-world agents and document processing. The technical report appeared on arXiv in February 2025, detailing the architectural changes and performance gains across the three model sizes.
Inside the model
The model uses a native dynamic-resolution Vision Transformer trained from scratch. Window attention reduces computational cost while keeping image details sharp. For video, it employs dynamic frame rate sampling and absolute time encoding to track events over long durations. The vision encoder aligns with the Qwen2.5 language model structure, using SwiGLU and RMSNorm. This design allows the system to process varying image sizes and video lengths without standard normalisation steps.
What it can do
It can pinpoint objects with bounding boxes or points and output coordinates in stable JSON format. The model extracts structured data from invoices, tables and diagrams. In benchmarks, the 72-billion parameter version scores highly on document question answering and chart interpretation. It also acts as a visual agent, capable of controlling computers and mobile devices by reasoning over screen content. Video comprehension includes identifying specific moments in clips lasting up to an hour.
How to use it
The weights are available openly through Hugging Face for all three sizes. The code integrates with the latest Hugging Face Transformers library. Users can run inference via Python scripts that handle image and video inputs. The team recommends building from source for full compatibility. Utilities are provided to manage visual data formats such as base64 strings and local file paths.
Questions people ask about Qwen2.5-VL
What sizes of Qwen2.5-VL are available?
The model family includes three variants with 3, 7 and 72 billion parameters. These sizes address different needs, from edge computing to high-performance server deployments.
Can it understand long videos?
Yes, it can comprehend videos lasting over an hour. It uses absolute time encoding to identify specific events and moments within the footage, rather than just summarising the whole clip.
Is the model open source?
The weights are openly available on Hugging Face. The code works with the Hugging Face Transformers library, allowing developers to download and run the models locally or on their own infrastructure.
Good reading and sources
There is no English Wikipedia article on Qwen2.5-VL yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2502.13923, huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.