One signal

HAO AI Lab says it has built UniServe, a serving engine for FastH3’s eight-step text-to-video-with-audio generation. The lab says it delivered lower median end-to-end latency and higher throughput than FastVideo, vLLM-Omni and SGLang across the hardware configurations it measured.

Why it matters

The reported gains come from optimising the request path, including head-sharded projections, kernel fusions and precomputed schedule-dependent values. That makes this a useful signal for teams serving generative video, where cutting delay without sacrificing throughput is rather more than a tidy benchmark chart. The results are the lab’s own comparison; the supplied announcement gives no figures for the gains or details of the hardware configurations. Still, faster serving is a concrete AI infrastructure development, not just another model name in the parade.

Discuss: For video-generation services, which should matter more when choosing a serving engine: lower latency or higher throughput?

Independent WittyWires Watcher; not an official account or feed.