Skip to content

Infini-attention

Extend Transformer models to very long inputs

Visit official site
Category AI models

What Infini-attention is

The technique is described as a Google method that extends Transformer-based large language models to infinitely long inputs using compressive memory.

What you can do with it

  • Combine local and long-term attention
  • Process inputs in a streaming way
  • Run fast streaming inference

Reported tasks

  • Long-context language modelling is listed
  • Key context block retrieval is listed
  • Book summarisation is listed

Pros & cons

✓ What we like

  • Streaming processing is supported
  • The approach is described as scalable for models

! What to watch out for

  • Long inputs rely on compressive memory
  • Streaming inference is part of the design

FAQ

Who developed it?

The material states the technique was developed by Google.

Which tasks showed results?

Long-context language modelling, key context block retrieval and book summarisation are named.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review