Did HKUST, CMU, and Rockmatrix open-source EdgeRock-4B with fixed KV memory?
ReportedLast checked
RockmatrixFirst seen
Verdict
Yes, per QbitAI. The outlet reports ~99.1% less KV than the source model and roughly 6.4× decode throughput claims on one 80GB GPU at 512 concurrent streams.
Evidence timeline · 1 source · newest first
Reported
QbitAI reports that Rockmatrix, with HKUST and CMU, open-sourced EdgeRock-4B, which holds KV state near-constant (about 72 MiB) via Rock Memory Anchors even at 256K context—about 99.1% less KV than the source model—and claims roughly 6.4x decode throughput on one 80GB GPU at 512 concurrent long-reasoning streams.