News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgvidFlamingo

Flamingo

A visual language model family that learns new image and video tasks from a few examples.

Flamingo is a family of visual language models described in an April 2022 paper by researchers at DeepMind and collaborators. It links pretrained vision and language models so a single system can handle interleaved images, video and text, and pick up new tasks from only a handful of labelled examples.

Share on WhatsApp

  • Family of visual language models first posted 29 April 2022
  • Learns new image and video tasks from a handful of examples
  • Accepts arbitrarily interleaved images, video and text
  • Bridges pretrained vision-only and language-only models
  • Published at NeurIPS 2022; 54-page paper with 27 authors

What it is

Flamingo is a family of visual language models, or VLMs, presented in a paper first posted on 29 April 2022 and later published at NeurIPS 2022. The work comes from a large author group led by Jean-Baptiste Alayrac, Jeff Donahue, Karen Simonyan and colleagues. The models combine a vision system with a language system so they can read a mix of pictures, video frames and words together. Their main purpose is few-shot learning: given just a handful of worked examples in a prompt, the model can take on a new task without any further training. The paper reports that one Flamingo model can handle many different image and video tasks this way.

How it came about

The Flamingo paper was submitted to arXiv on 29 April 2022, with a revised version following on 15 November 2022. Twenty-seven authors are listed, spanning vision, language and reinforcement learning research. The work appeared in the proceedings of the Neural Information Processing Systems conference in 2022. The paper is 54 pages long and sits in the computer vision, artificial intelligence and machine learning categories.

Inside the model

Flamingo connects two kinds of pretrained model: one trained only on images, and one trained only on text. The authors describe architectural changes that let the combined system accept sequences where pictures and words are mixed in any order, and let it take either still images or video as input. Because the design tolerates that mixture, the models can be trained on large collections of web data where text and images are interleaved. That training is what gives them their few-shot ability. The paper does not state parameter counts, layer counts or training compute in the abstract.

What it can do

Flamingo is evaluated on open-ended tasks such as visual question answering, where it is shown a question about an image and must answer, and captioning, where it describes a scene or event. It is also tested on closed tasks such as multiple-choice visual question answering. The authors report that a single Flamingo model reaches a new state of the art through few-shot prompting across this range of tasks, and on many benchmarks it beats models fine-tuned on thousands of times more task-specific data.

How to use it

The paper is publicly available on arXiv and was published at NeurIPS 2022. The abstract does not describe released weights, an API or a licence.

Why it matters

Flamingo set out a way to reuse strong single-modality models inside one multimodal system, and showed that in-context examples could replace task-specific fine-tuning for many image and video problems. Its prompting-based evaluation style, where one model is tested across many tasks with only a few examples each, became a common way to report visual language model results.

Questions people ask about Flamingo

What is Flamingo?

Flamingo is a family of visual language models introduced in an April 2022 paper. It combines a pretrained vision model with a pretrained language model so one system can work with mixtures of images, video and text, and can take on new tasks from only a few examples given in the prompt.

Who created it?

A group of 27 researchers, with Jean-Baptiste Alayrac, Jeff Donahue, Karen Simonyan and Andrew Zisserman among the listed authors. The paper was published in the proceedings of NeurIPS 2022.

What does few-shot learning mean here?

It means the model is given a small number of worked examples in the prompt and adapts to the task immediately, with no further training. The authors report this works across question answering, captioning and multiple-choice tasks on images and video.

What inputs can it handle?

Sequences where text and images are interleaved in any order, plus video as well as still pictures. That flexibility is what allowed it to be trained on large web collections of mixed text and images.

How large are the models?

The abstract does not give parameter counts, layer numbers or training compute. It describes Flamingo as a family of models rather than a single size.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Flamingo yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2204.14198. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.