Rock Matrix open-sources EdgeRock-4B with 72 MiB KV cache

Cache engineering at this scale makes long-context reasoning practical on far smaller hardware.

EdgeRock-4B, built from Qwen3.5-4B with HKUST and CMU, keeps a fixed 72 MiB KV cache at 256K context, down from 8 GiB, serves 6.42x the throughput while keeping most reasoning scores.

Name

EdgeRock-4B

Params

4B

Context

256K

Domain

general

Read it at Pandaily →

Coverage: 1 sources, in order

1. 27 Oct 02:52 Pandaily Read the article →

← Front page