Flamingo is a family of Visual Language Models (VLMs) designed for few-shot learning, capable of adapting to new tasks with minimal annotated examples. The model bridges powerful pre-trained vision-only and language-only models, handles sequences of arbitrarily interleaved visual and textual data, and seamlessly processes images or videos. Flamingo models are trained on large-scale multimodal web data, enabling in-context few-shot learning. They outperform models fine-tuned on thousands of task-specific examples on various benchmarks, including visual question-answering, captioning, and multiple-choice tasks. Flamingo's architecture includes a Perceiver Resampler to reduce computational complexity and a GATED XATTN-DENSE layer to condition language models on visual inputs. The model is trained on a mixture of web-based datasets, including interleaved text and image data, and can be fine-tuned for additional tasks. Flamingo demonstrates strong performance across 16 multimodal tasks, with a single model achieving state-of-the-art results with few-shot learning. The model's flexibility allows it to process varying numbers of images and videos, and it shows robust performance even with limited training data. Flamingo's design enables efficient adaptation to new tasks without extensive fine-tuning, making it a significant advancement in visual language modeling.Flamingo is a family of Visual Language Models (VLMs) designed for few-shot learning, capable of adapting to new tasks with minimal annotated examples. The model bridges powerful pre-trained vision-only and language-only models, handles sequences of arbitrarily interleaved visual and textual data, and seamlessly processes images or videos. Flamingo models are trained on large-scale multimodal web data, enabling in-context few-shot learning. They outperform models fine-tuned on thousands of task-specific examples on various benchmarks, including visual question-answering, captioning, and multiple-choice tasks. Flamingo's architecture includes a Perceiver Resampler to reduce computational complexity and a GATED XATTN-DENSE layer to condition language models on visual inputs. The model is trained on a mixture of web-based datasets, including interleaved text and image data, and can be fine-tuned for additional tasks. Flamingo demonstrates strong performance across 16 multimodal tasks, with a single model achieving state-of-the-art results with few-shot learning. The model's flexibility allows it to process varying numbers of images and videos, and it shows robust performance even with limited training data. Flamingo's design enables efficient adaptation to new tasks without extensive fine-tuning, making it a significant advancement in visual language modeling.