The paper introduces Florence, a new computer vision foundation model designed to generalize well across a wide range of tasks with minimal customization. Florence is trained on a large-scale, diverse dataset of 900 million image-text pairs, enabling it to handle tasks such as classification, retrieval, object detection, VQA, image captioning, video retrieval, and action recognition. The model's architecture includes a hierarchical Vision Transformer (CoSwin) and a language encoder, allowing it to learn representations from coarse to fine, static to dynamic, and single modality to multiple modalities. Florence demonstrates superior performance in zero-shot, few-shot, and zero-shot transfer learning, achieving state-of-the-art results on various benchmarks, including ImageNet-1K, COCO, VQA, and Kinetics-600. The paper also details the model's training infrastructure, which includes techniques to reduce memory consumption and increase training efficiency. Overall, Florence represents a significant advancement in the field of computer vision, offering a versatile and powerful foundation for a wide range of vision tasks.The paper introduces Florence, a new computer vision foundation model designed to generalize well across a wide range of tasks with minimal customization. Florence is trained on a large-scale, diverse dataset of 900 million image-text pairs, enabling it to handle tasks such as classification, retrieval, object detection, VQA, image captioning, video retrieval, and action recognition. The model's architecture includes a hierarchical Vision Transformer (CoSwin) and a language encoder, allowing it to learn representations from coarse to fine, static to dynamic, and single modality to multiple modalities. Florence demonstrates superior performance in zero-shot, few-shot, and zero-shot transfer learning, achieving state-of-the-art results on various benchmarks, including ImageNet-1K, COCO, VQA, and Kinetics-600. The paper also details the model's training infrastructure, which includes techniques to reduce memory consumption and increase training efficiency. Overall, Florence represents a significant advancement in the field of computer vision, offering a versatile and powerful foundation for a wide range of vision tasks.