The QVQ-72B AI model outperformed OpenAI o1 on the Math Vista benchmark.
The QVQ-72B AI model can be accessed from Hugging Face
Photo Credit: Unsplash/Markus Winkler
Alibaba's Qwen research team has released another open-source artificial intelligence (AI) model in preview. Dubbed QVQ-72B, it is a vision-based reasoning model that can analyse visual information from images and understand the context behind them. The tech giant has also shared benchmark scores of the AI model and highlighted that on one specific test, it was able to outperform OpenAI's o1 model. Notably, Alibaba has released several open-source AI models recently, including the QwQ-32B and Marco-o1 reasoning-focused large language models (LLMs).
In a Hugging Face listing, Alibaba's Qwen team detailed the new open-source AI model. Calling it an experimental research model, the researchers highlighted that the QVQ-72B comes with enhanced visual reasoning capabilities. Interestingly, these are two separate branches of performance, that the researchers have combined in this model.
Vision-based AI models are plenty. These include an image encoder and can analyse the visual information and context behind them. Similarly, reasoning-focused models such as o1 and QwQ-32B come with test-time compute scaling capabilities that allow them to increase the processing time for the model. This enables the model to break down the problem, solve it in a step-by-step manner, assess the output and correct it against a verifier.
With QVQ-72B's preview model, Alibaba has combined these two functionalities. It can now analyse information from images and answer complex queries by using reasoning-focused structures. The team highlights that it has significantly improved the performance of the model.
Sharing evals from internal testing, the researchers claimed that the QVQ-72B was able to score 71.4 percent in the MathVista (mini) benchmark, outperforming the o1 model (71.0). It is also said to score 70.3 percent on the Multimodal Massive Multi-task Understanding (MMMU) benchmark.
Despite the improved performance, there are several limitations, as is the case with most experimental models. The Qwen team stated that the AI model occasionally mixes different languages or unexpectedly switches between them. The code-switching issue is also prominent in the model. Additionally, the model is prone to getting caught in recursive reasoning loops, affecting the final output.
Catch the latest from the Consumer Electronics Show on Gadgets 360, at our CES 2026 hub.