Alibaba Cloud, the renowned cloud computing division of Alibaba Group Ltd., has made a significant stride in artificial intelligence with the announcement of its latest experimental model, the QVQ-72B-Preview. This open-source AI model introduces advanced capabilities in visual reasoning, promising to enhance the way machines interpret and draw conclusions from images.
During its announcement on Wednesday, Alibaba shared that initial benchmarks positioned the QVQ-72B-Preview as a leading model in visual reasoning, capable of methodically solving problems in a manner akin to models such as OpenAI’s o1 and Google LLC’s Gemini Flash.
The QVQ model is a continuation of the Qwen family, building on the advanced analysis and reasoning capabilities of its predecessor, the Qwen2-VL-72B, renowned for its video analysis prowess. With QVQ, Alibaba has introduced an AI that mirrors the cognitive process of a master physicist, contemplating complex problems and deducing solutions with remarkable precision.
"Imagine an AI that can look at a complex physics problem, and methodically reason its way to a solution with the confidence of a master physicist. This vision inspired us to create QVQ – an open-weight model for multimodal reasoning," expressed the Qwen team.
Users engage with the QVQ-72B by submitting an image coupled with a prompt. The model intricately analyzes the image, offering a comprehensive step-by-step response. It outlines its thought process, enabling users to understand its method of reasoning. For instance, when tasked with counting fish in an aquarium, the model meticulously identifies and counts the fish, verifying its observations through multiple perspectives to ensure accuracy.
The QVQ-72B-Preview underwent rigorous testing across four key datasets, including MMMU, MathVista, MathVision, and OlympiadBench, consistently delivering results that paralleled or surpassed other high-performance closed-source models. Notably, within the MMMU benchmark, it achieved a noteworthy score of 70.3, closely mirroring the results of Claude 3.5 Sonnet from Anthropic PBC.
Despite its promising capabilities, the QVQ-72B remains in the experimental phase, with several areas identified for further development. Presently, the model faces challenges in language mixing during response formulation and exhibits verbose tendencies. Enhancements in safety measures will be a prerequisite before a broader release.
Alibaba’s QVQ-72B is available under the open-source Qwen license on platforms like GitHub and Hugging Face, inviting developers and researchers to refine and expand its functionalities. The Qwen team regards this model as a milestone in the quest towards achieving artificial general intelligence (AGI), aiming to integrate vision-based cognition seamlessly with reasoning capabilities.
Alibaba's cutting-edge QVQ-72B model represents a bold leap towards achieving integrated AI vision and reasoning, setting the stage for future advancements towards AGI. Jengu.ai, renowned for its expertise in automation, AI, and process mapping, continues to deliver insightful content, positioning its audience at the forefront of these revolutionary technological advancements.
```