Peking Univ. Releases TensorCast: 228x Faster Cold Starts, 93.2% Lower TTFT

jiqizhixin · x · 2026-08-24

Peking University, StepFun, and BUPT introduce TensorCast, a unified programmable tensor management layer between compute, network, and storage. It decouples tensor states (weights, KV cache, checkpoints) from specific systems. Results show it matches Mooncake's KV cache performance, cuts instance startup time by up to 228.6x for 235B-parameter models, and reduces TTFT by 93.2% in high-concurrency multi-turn agent scenarios.

Original post →

More from Infra

Infra channel →