QVQ-Max: Thinking with Evidence

Going beyond the limitations of QVQ-72B-Preview released last December, the first version of the visual reasoning model QVQ-Max has been officially launched. This model goes beyond simply understanding the content of images and videos, demonstrating reasoning capabilities across various domains from math problems to programming and artistic creation.

According to MathVision benchmark results, accuracy continued to improve as the length of the model's thinking process was adjusted, proving its strong potential.

The core capabilities of QVQ-Max can be summarized into three areas:

• Detailed Observation: Quickly identifies key elements and subtle details in complex charts or everyday photos.
• Deep Reasoning: Combines visual information with background knowledge to solve geometry problems or predict the next scene in a video.
• Flexible Application: Capable of creative tasks such as assisting with illustration design, generating video scripts, and creating role-play content.

Its range of applications is also broad. At work, it helps with data analysis and code writing, while for students it provides math and physics problem-solving and concept explanations. In daily life, it can offer practical help such as outfit recommendations or recipe guides.

In future development stages, the plan is to focus on improving observation accuracy through grounding technology, implementing Visual Agent functionality that operates smartphones or computers, and more advanced interaction methods.