Flamingo

Flamingo is a vision-language model DeepMind published in April 2022, built to handle images, videos, and text together and answer open-ended questions about them after seeing only a handful of examples. It starts from Chinchilla, DeepMind’s 70B parameter language model, and adds vision encoders and cross-attention layers that let it condition text generation on visual input, bringing the full model to roughly 80B parameters. The key idea was few-shot learning for vision-language tasks: rather than fine-tuning on each new dataset, Flamingo could be shown a small number of examples in its prompt and generalize from there, similar to how GPT-3 did for text. At release it beat every prior few-shot approach on 16 vision-language benchmarks, including captioning, visual question answering, and visual dialogue. DeepMind never released Flamingo as a product or public API; it stayed a research model, but its architecture influenced later multimodal systems, including open reimplementations like OpenFlamingo and design choices in Google’s subsequent multimodal models.