[KongchangAI]
ModelKVZip+

KVZip+

A KV cache compression method based on budget matching. It first estimates an importance score for each cache entry, then retains or merges entries under a given budget. It requires no additional training, but suffers from severe information loss at high compression rates.

Timeline (last 90 days)

Oct 15

Latent Compaction achieves 96.7% accuracy on the RULER benchmark at a 32x compression rate, while KVZip+ achieves only 40.0%.

Unverified50%

All Facts (1)

Source Articles