Rapidly acquiring three-dimensional (3D) building data, including geometric attributes like rooftop, height, and structure, as well as indicative attributes like function, quality, and age, is essential for accurate urban analysis, simulations, and policy updates. Existing large-scale building datasets lack accuracy, extensibility and indicative attributes. This paper presents a geospatial artificial intelligence (GeoAI) framework for large-scale building modeling, introducing the first Multi-Attribute Building dataset (CMAB) in China at a national scale. The dataset covers 3,667 natural cities with a total rooftop area of 21.3 billion square meters with an F1-Score of 89.93% in rooftop extraction through the OCRNet. We trained bootstrap aggregated XGBoost models with city administrative classifications, incorporating building features such as morphology, location, and function. Using multi-source data, including billions of high-resolution Google Earth imagery and 60 million street view images (SVI), we generated rooftop, height, function, age, and quality attributes for each building. Accuracy was validated through model benchmarks, existing similar products, and manual SVI validation. The results support urban planning and sustainable development.