Qwen2.5-VL Technical Report

Qwen2.5-VL is the latest flagship vision-language model of the Qwen series, designed to understand and interact with the world through enhanced visual recognition, precise object localization using bounding boxes or points, robust document parsing, structured data extraction from invoices, forms and tables, and long-video comprehension. It introduces dynamic resolution processing and absolute time encoding so it can natively perceive spatial scales and temporal dynamics, processing images of varying sizes and videos up to hours long with second-level event localization, and trains a native dynamic-resolution Vision Transformer with window attention to reduce computational overhead.

Beyond static image and document understanding, the model serves as an interactive visual agent capable of reasoning, tool usage, and task execution such as operating computers and mobile devices, and it is offered in three sizes spanning edge AI to high-performance computing. The flagship Qwen2.5-VL-72B model matches state-of-the-art systems like GPT-4o and Claude 3.5 Sonnet, particularly excelling in document and diagram understanding, while preserving the core language competencies of the underlying Qwen2.5 LLM.