QVQ-Max explored whether giving a visual AI more room to reason could improve answers that depend on diagrams. Qwen reported that increasing the permitted reasoning length improved accuracy on MathVision, a test of mathematical problems involving visual information. The model accepts images and videos and combines their contents with reasoning to produce answers. The March 2025 release followed the QVQ-72B research preview. Its reported result supports longer reasoning as a useful control to investigate for visual problem solving; computer operation and tool-based verification remained future development areas.
Terms in this piece · Glossary
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
Key quotes
“Today, we are officially releasing the first version of QVQ-Max, our visual reasoning model.”
