Introducing Qwen-VL

The Qwen-VL series, which expands Qwen's multimodal capabilities, has been upgraded into two enhanced versions: Qwen-VL-Plus and Qwen-VL-Max. This update significantly improves image-related reasoning capabilities as well as the ability to recognize, extract, and analyze text and detailed information within images.

In particular, it supports high-resolution images over 1 million pixels and various aspect ratios.

• qwen-vl-plus: A model with enhanced detail recognition and text recognition capabilities, supporting ultra-high-resolution images and arbitrary aspect ratios.
• qwen-vl-max: The most powerful model, with further improved visual reasoning and instruction-following capabilities, delivering optimal performance on complex tasks.

In terms of performance, these models match Gemini Ultra and GPT-4V across multiple multimodal tasks, significantly surpassing the records of existing open-source models. In particular, on Chinese question-answering and text comprehension tasks, they outperform GPT-4V and Gemini.

In real-world use, they have also demonstrated outstanding problem-solving abilities in conversation, identifying celebrities and landmarks, text generation, and interpreting visual content.