Qwen2.5-VL

A local vision-language model analyzes images, extracts document data and locates video events, with Apache 2.0 weights and Transformers support.

Qwen2.5-VL is a vision-language model for developers building visual assistants, document processing applications and screen-based agents. It answers questions about images and video and can return structured data for software to use.

Document extraction: Read text, charts and page layouts, then extract structured content from scanned invoices, forms and tables.
Image analysis and comparison: Identify objects and interpret graphics, or examine multiple images in one query to compare their contents.
Object locations: Return bounding boxes or points for objects, with coordinates and attributes in JSON output.
Video event search: Analyze videos longer than an hour and identify the segments associated with a requested event, linking its answers to specific moments.
Visual agent reasoning: Interpret computer and phone screens and direct tools for interaction within an application that provides those tools.

The instruction-tuned Qwen2.5-VL-7B-Instruct weights use the Apache 2.0 license and come in Safetensors format. The model works with Hugging Face Transformers and supports local image and video files, so local inference doesn't require sending those inputs to a cloud service.

CUDA GPU inference is supported. Adjustable image resolution lets applications trade visual detail against computation and memory use; FlashAttention 2 can reduce memory use for multiple-image and video workloads.