Qwen-VL-Plus and Qwen-VL-Max Release

Qwen has introduced Qwen-VL-Plus and Qwen-VL-Max, two enhanced versions of the Qwen-VL series. These models significantly improve image-related reasoning, detailed recognition, and the ability to process high-definition images with resolutions exceeding one million pixels and arbitrary aspect ratios.

Model Comparison: Qwen-VL-Plus vs. Qwen-VL-Max

Qwen provides two distinct versions of the enhanced visual language model to balance performance and capability:

  • Qwen-VL-Plus: An enhanced large visual language model designed for detailed recognition and text recognition abilities, supporting ultra-high pixel resolutions and arbitrary aspect ratios.
  • Qwen-VL-Max: The most capable model in the series, offering further improvements in visual reasoning and instruction-following capabilities for higher-level cognitive understanding and complex tasks.

Performance Benchmarks

Qwen-VL-Plus and Qwen-VL-Max perform on par with Gemini Ultra and GPT-4V across multiple text-image multimodal tasks. Notably, Qwen-VL-Max outperforms both GPT-4V and Gemini in Chinese question answering and Chinese text comprehension.

Model DocVQA ChartQA AI2D TextVQA MMMU MathVista MM-Bench-CN
Other Best Open-source LVLM 81.6% 68.4% 73.7% 76.1% 45.9% 36.7% 72.4%
Gemini Pro 88.1% 74.1% 73.9% 74.6% 47.9% 45.2% 74.3%
Gemini Ultra 90.9% 80.8% 79.5% 82.3% 59.4% 53.0% -
GPT-4V 88.4% 78.5% 78.2% 78.0% 56.8% 49.9% 73.9%
Qwen-VL-Plus 91.4% 78.1% 75.9% 78.9% 45.2% 43.3% 68.0%
Qwen-VL-Max 93.1% 79.8% 79.3% 79.5% 51.4% 50.0% 75.1%

Key Technical Capabilities

High-Resolution Image Processing

Qwen-VL-Plus and Qwen-VL-Max support images with resolutions above one million pixels and images of various aspect ratios, allowing for more precise recognition of details within images and texts.

Visual Reasoning and Problem Solving

Beyond simple description, the models can interpret complex representations such as flowcharts, diagrams, and symbolic systems. This includes mathematical problem-solving and the analysis of charts and graphs.

Visual Agent and Localization

The models possess the ability to pinpoint and query specific elements within an image, such as highlighting specific objects (e.g., a red car) and making decisions based on the same context of a scene.

Text Information Recognition

Qwen-VL-Plus/Max exhibit significantly improved recognition of Chinese and English text. They can efficiently extract information from documents and tables and reformat it into custom output formats, such as JSON, for dense text processing.

Availability

Qwen-VL-Plus and Qwen-VL-Max are available through:

  • Huggingface Spaces: Demos for both Plus and Max versions.
  • QianWen Web Portal: Image Understanding mode for Qwen-VL-Max.
  • Dashscope APIs: API access for both models.

Sources