Flamingo

Flamingo, introduced by DeepMind in April 2022, was the first visual language model family with GPT-3-style few-shot learning across modalities. It bridges a frozen vision encoder (NFNet-F6) and a frozen language model (Chinchilla at 1.4B, 7B, and 70B scales) with two small trainable modules: a Perceiver Resampler that compresses visual features into a fixed bank of latent vectors, and gated cross-attention layers interleaved into the frozen LM.

Crucially, Flamingo accepts arbitrarily interleaved sequences of images, video, and text, so in-context few-shot prompting works across modalities: show it two captioned examples and a third image, and it captions the third. The 80B model set new state of the art on 16 multimodal benchmarks with few-shot prompting alone. It was never publicly released, so this is a research-landmark entry: its Perceiver Resampler and frozen-backbone recipe became the template that BLIP-2, LLaVA, and most later VLMs inherited.