Qwen2.5-VL Released

Qwen has released Qwen2.5-VL, its next-generation flagship vision-language model (VLM), marking a significant leap from its predecessor, Qwen2-VL. This release comes in three sizes—3B, 7B, and 72B—available as both Base and Instruct models.

Key features include:

• Visual understanding: Proficient not only at recognizing common objects but also at analyzing text, charts, icons, and layouts.
• Agentic capabilities: Functions as a visual agent capable of using computers and smartphones, dynamically directing tools and reasoning.
• Long-video comprehension: Understands videos over 1 hour long and can accurately pinpoint specific events.
• Precise localization and structured output: Accurately localizes objects using bounding boxes or points, and can convert data such as invoices or tables into structured data in JSON format.

In terms of performance, Qwen2.5-VL-72B-Instruct delivers competitive results against SOTA models across various benchmarks, including math, document understanding, and video understanding. It shows particular strength in document and chart comprehension, and can perform as a visual agent without additional fine-tuning.

For small and medium-sized models, Qwen2.5-VL-7B-Instruct outperforms GPT-4o-mini on multiple tasks. The 3B model, designed for edge AI, achieves better performance than the previous Qwen2-VL 7B model.