Alibaba releases Qwen2.5-VL as a visual agent model

Background

Qwen2.5-VL is a vision-language model from Alibaba's Qwen team. A model of this kind takes images or video together with text and answers in text, so it can read a chart, describe a scene or find an object in a picture. The release extended that ability in three directions at once: acting as an agent that operates software, following long videos, and producing precise coordinates that other programs can use.

What happened

In January 2025 Alibaba released Qwen2.5-VL in four sizes: 3B, 7B, 32B and 72B. Every size except the 72B model was published under the Apache 2.0 license.^[2]

The model was described as a visual agent. It can reason about what it sees and direct tools, including tasks on a computer or a phone.^[1]

Key details

Long video is one of its strengths. Qwen2.5-VL can understand videos longer than one hour and identify the segments that matter for a question.^[1]

Its outputs are structured. The model can locate objects with bounding boxes or points, and it produces stable JSON for coordinates, attributes and document content such as invoices and forms.^[1]

The published scores for the 72B model were 70.2 on MMMU, 74.8 on MathVista-MINI, 96.4 on DocVQA and 89.5 on ChartQA. On VideoMME it scored 73.3 without subtitles and 79.1 with subtitles.^[1] On agent benchmarks it scored 87.1 on ScreenSpot and 43.6 on ScreenSpot Pro.^[1]

Some benchmark names in the scores need explaining. MMMU tests understanding across academic subjects using images together with text, MathVista tests mathematical reasoning with visual content, DocVQA tests question answering on document images, and ChartQA tests reading and reasoning about charts. ScreenSpot tests whether a model can locate elements on a screen, and ScreenSpot Pro is a more demanding version of that task.^[1]

Why it matters

The release placed several visual abilities in one model family: reading screens and documents, following long video, and locating objects. Because the outputs are structured, another program can take them as input without parsing free text. The smaller sizes were released under a permissive license, which allows them to be used and modified under its terms.^[2]