SYLIN

Nari Qwen3-TTS

by Nari Labs · Runtime · Voice

Open-source optimized serving of Qwen3-TTS 1.7B CustomVoice with HTTP streaming and WebSocket text input; needs an H100

Evidence

  1. Aug 21
    Claims 34 ms p95 time-to-first-audio at 10 requests/s on one H100; sub-80 ms p95 at 20 requests/s
  2. Aug 21
    Authors say vLLM-Omni and SGLang-Omni are too slow for production and can break realtime playback at low latency

Alternatives

Works with