QVQ: Looking at the World with Wise Sight

The Qwen team developed QVQ, an open-weight multimodal reasoning model based on Qwen2-VL-72B, to maximize visual understanding and complex problem-solving ability. QVQ is designed to combine linguistic thinking with visual memory, enabling it to reason step by step through complex physics problems and the like.

In terms of performance, QVQ scored 70.3 on the MMMU benchmark, significantly outperforming the existing Qwen2-VL-72B-Instruct. It also showed excellent results on the following math- and science-focused benchmarks, effectively narrowing the gap with the leading o1 model:

• MathVista: Evaluates logical, algebraic, and scientific visual reasoning
• MathVision: A high-quality multimodal math reasoning test set
• OlympiadBench: Olympiad-level math and physics problems

However, as an experimental research model, it also has some limitations. Language mixing, where languages unexpectedly mix together during the answering process, or recursive reasoning problems, where it falls into circular logic without reaching a conclusion, can occur. In addition, during multi-step visual reasoning, it may lose focus on the image content and cause hallucination, so caution is needed.