Skip to content
gentic.news — AI News Intelligence Platform

Technique · inference

StreamingLLM (Attention Sinks)

A sliding-window attention pattern with preserved initial tokens ("sinks") that enables indefinite streaming generation without quality collapse.

Origin: MIT, 2023-09Read origin paper →Also known as: StreamingLLM, Attention Sinks
0
Products deploying
Avg research → prod
First commercial deploy

Deployment timeline

No verified deployments yet in our tracked product set.