Qwen2.5-VL

2025-01-26 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen2.5-VL in three sizes (3B, 7B, 72B): an open-weights vision-language model able to act as a visual agent, understand hour-long videos and produce structured outputs.

### Context Qwen2-VL (August 2024) had laid the foundations of Qwen's vision-language family. In January 2025, the comparison is made against GPT-4o, Claude 3.5 Sonnet and Gemini 2.0 Flash, which the 72B model card uses as references. ### What the model brings - Broader visual recognition (objects, places, characters from films and series, products) and analysis of text, charts, icons and layouts. - Visual agent role: the model reasons and drives tools, with computer and phone use capabilities. - Understanding of videos longer than one hour, with localization of the relevant segments. - Localization by bounding boxes or points, with stable JSON outputs for coordinates and attributes. - Structured outputs for invoices, forms and tables, useful in finance and commerce. ### Architecture and specifications Three sizes: 3B, 7B and 72B, in base and Instruct versions. Licenses declared on Hugging Face: Apache 2.0 for the 7B, “Qwen Research” license for the 3B and “Qwen” license for the 72B (commercial use to be checked in the license text). Modalities: text, images, video. Context length, languages and knowledge cutoff: not verified in the sources consulted. Weights on Hugging Face and ModelScope. ### Cost and access Downloadable weights; demo on Qwen Chat. Vendor API pricing and required hardware: not verified for this entry. ### Benchmarks Excerpt from the Hugging Face model card of the 72B (scores published by the vendor, comparison with closed models). | Benchmark | GPT-4o | Claude 3.5 Sonnet | Gemini 2.0 Flash | Qwen2.5-VL-72B | |---|---|---|---|---| | MMMU (val) | 70.3 | 70.4 | 70.7 | 70.2 | | MathVista (mini) | 63.8 | 65.4 | 73.1 | 74.8 | | MathVision | 30.4 | 38.3 | 41.3 | 38.1 | | DocVQA (val) | 91.1 | 95.2 | 92.1 | 96.4 | | ChartQA | 86.7 | 90.8 | 85.2 | 89.5 | | OCRBench | 736 | 788 | not published | 885 | | VideoMME without subtitles | 71.9 | not published | not published | 73.3 | Source: Hugging Face model card for Qwen2.5-VL-72B-Instruct, vendor measurements; the video table also compares Gemini 1.5 Pro (75.0 on VideoMME without subtitles). ### Versus closed models On document understanding, the 72B scores 96.4 on DocVQA versus 95.2 for Claude 3.5 Sonnet and 91.1 for GPT-4o, and 885 on OCRBench versus 788 and 736. On MMMU, it is on par (70.2 versus 70.3 to 70.7). On the other hand, on MathVision, Gemini 2.0 Flash (41.3) and Claude 3.5 Sonnet (38.3) exceed or match the open model (38.1). On computer-agent tasks, the vendor publishes an OSWorld score of 8.83 versus 14.90 for Claude, so it is behind. The vendor also claims that the 7B surpasses GPT-4o-mini on several tasks. The price gap with closed APIs is not established here; the main benefit is self-hosting for sensitive documents. ### Key takeaways Qwen2.5-VL puts an open model on par with the best closed models of January 2025 on reading documents and charts. Licenses vary by size: check the 72B's license before any commercial use.

Tags : LLM, Open Source, Alibaba, Qwen, Vision, Apache 2.0, Multimodal