ACCT enables KV cache transfer between models of different scales
November 2, 2026
ACCT allows for asymmetric context transfer by transforming KV caches from a source model into sparse states consumable by a target model within the same family. This method decouples context computation from inference, reducing redundant processing costs in heterogeneous model deployments.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference latency and compute costs in multi-model pipelines by reusing KV caches across different model sizes.
●
researcherThis introduces a method for cross-model context reuse via sparse KV state transformation.