METASTONE launches TokenFleet, a distributed scheduling framework for inference clusters
AIMETASTONE launched TokenFleet, a distributed scheduling framework for GPU inference clusters, at the 2026 CCF National High Performance Computing Conference. Its TokenPilot module reports a 32.4% cut in P95 time to first token in tests on a GLM5.3 workload.
AIWhy it matters.
The article explains how request routing, cross-datacenter capacity sharing, and workload admission work together, and defines goodput as the metric it optimizes.
Loading source posts…
Source details
QbitAI · 思邈 · qbitai.comPublished · added here
Original title: 首Token延迟降低32.4%！是石科技集群调度大招，国产推理再提效