Qwen2.5-VL

What it is

Qwen2.5-VL is a vision-language model family from Alibaba's Qwen team. It takes images and video together with text, and it produces text answers. It was released in January 2025 in 3B, 7B, 32B and 72B sizes, and all sizes except the 72B use Apache 2.0.^[2] Qwen2.5-VL-32B-Instruct was added on 24 March 2025, also under Apache 2.0.^[2]

Origins

Qwen2.5-VL is the successor to Qwen2-VL. Its architecture extends dynamic resolution and frame-rate sampling to time, aligns mRoPE to absolute time, and uses a vision encoder with window attention, together with SwiGLU and RMSNorm components.^[1]

Main features

The model reads text, charts, icons, graphics and layouts, as well as common objects.^[1] It can act as a visual agent that reasons and directs tools, including computer and phone use.^[1] It can also understand videos over one hour long and pinpoint the relevant event segments.^[1]

It localises objects with bounding boxes or points, and it outputs stable JSON for coordinates, attributes and documents such as invoices and forms.^[1] Structured output of this kind lets a program read the model's answers directly, without parsing free text.

For benchmarks, the Qwen2.5-VL-72B model scored 70.2 on MMMU, 74.8 on MathVista-MINI, 96.4 on DocVQA and 89.5 on ChartQA. On the video benchmark VideoMME it scored 73.3 without subtitles and 79.1 with them.^[1] On agent benchmarks it scored 87.1 on ScreenSpot and 43.6 on ScreenSpot Pro.^[1]

Several benchmarks named above test different skills. MMMU tests multidisciplinary understanding of images, DocVQA tests question answering on document images, ChartQA tests reading and reasoning about charts, and VideoMME tests video understanding. ScreenSpot tests whether a model can locate an element on a graphical screen from a description. The scores are the team's reported figures, and they show performance on those tests rather than on every task.

Its place in development

Qwen2.5-VL followed Qwen2-VL.^[2] It was succeeded by Qwen3-VL in September 2025.^[2]