Nari Labs open-sources a sub-50ms Qwen3-TTS voice server
A community build brings the same open TTS to a consumer 4090; both self-host with no per-minute bill — plus a Rust LSP that fits in 100MB.
Copy markdown
Sub-50ms p95 on a single H100
Nari Labs open-sourced (Apache-2.0) a serving stack for Qwen3-TTS 1.7B CustomVoice that holds sub-50ms p95 time-to-first-audio at 10 requests/second on one H100 — and sub-80ms at 20 RPS. It ships HTTP and WebSocket APIs via Docker or uv, ready to wire into a live voice loop.
Or run it on a consumer 4090
andimarafioti's faster-qwen3-tts (MIT) swaps dynamic caching for CUDA-graph capture to reach 156ms streaming TTFA and 4.78x real-time on a single RTX 4090 with the 0.6B model — a 5.8x throughput jump over stock. pip-installable, with voice cloning, CustomVoice, and VoiceDesign.
Build this weekend: a voice agent with no meter running
Both stacks self-host Alibaba's open-weights Qwen3-TTS, so a full mic-to-LLM-to-speech loop runs on your own GPU with no per-character API bill. Clone a voice from a few seconds of reference audio, or design one from a text prompt.
Elsewhere: a Rust LSP that fits in 100MB
Rust Glancer, a new open-source alternative to rust-analyzer, holds memory under 100MB and re-indexes instantly on restart — headroom that matters when a coding agent is already eating your RAM. It matched rust-analyzer's index time on an 8GB M1, and the VS Code extension is live.