Qwen2.5-VL: Alibaba's open VLM that beats GPT-4o on DocVQA
In one sentence Alibaba releases Qwen2.5-VL in three sizes (3B, 7B, 72B), base and instruct: document reading with structured output, videos over an hour long and visual-agent capabilities. The 72B scores 96.4 on DocVQA versus 91.1 for GPT-4o.
Qwen2.5-VL is Alibaba's new family of vision-language models, published on January 26, 2025 in three sizes (3, 7 and 72 billion parameters), with weights downloadable from Hugging Face and ModelScope. It can read documents, invoices, charts and forms and return their content in structured form; it understands videos longer than an hour and can point to the second at which an event happens; it can act as a visual agent, operating a computer or a phone from screenshots.
According to the numbers published by Alibaba, the 72B is competitive with the best closed models and beats them on some document tests: on DocVQA it scores 96.4 versus 91.1 for GPT-4o. The 7B outperforms GPT-4o-mini on several tasks and the 3B beats the previous 7B Qwen2-VL.
Companies
Alibaba
Tools
Qwen2.5-VL, Qwen2.5-VL-72B, Qwen2.5-VL-7B, Qwen2.5-VL-3B
Tags
Sources