ElevenLabs V3 charges $100 per million characters. Cartesia Sonic 3.5 charges $49. Nari Labs just open-sourced an inference stack that serves the same job for about $2 — 25-50x cheaper, with better latency. HN gave it 118 points in a day.
Not a new model, just brutal inference engineering
This is a serving implementation of Qwen3-TTS 1.7B CustomVoice, Alibaba’s open-weights voice model. Nari rewrote the pipeline: one scheduler runs all three components (Talker, Code Predictor, Codec) as independently schedulable tasks instead of a fixed sequence, requests still waiting for their first audio chunk get priority over ones already streaming, plus CUDA graphs and a state-cached codec. On a single H100: 10 requests per second, 34ms p95 time-to-first-audio, 630 characters per second at full load.
Self-host it as your own TTS API
The whole thing ships as open source with streaming and non-streaming HTTP endpoints plus WebSocket text input — deploy one H100 and you have a voice API for real-time agents, dubbing, or call automation. The team previously built Dia, the dialogue TTS model with 2M+ downloads.
Open weights plus inference engineering is now undercutting commercial voice APIs by 50x. That pricing model doesn’t survive this.
You Might Also Like
- Mistral Voxtral tts Scores 63 Listener Preference Over Elevenlabs and the Weights are Free
- Embarrassingly Simple Self Distillation ssd Boosts Qwen3 30b Code Scores by 30 no Teachers no rl no Tricks
- Palmer Luckey Wants Your Python Code to fly a Drone at 100 mph Anduril ai Grand Prix is now Open
- Roborock Saros 20 Sonic Launches Globally With a 7 98cm Chassis and an Open api Your Home Assistant Agent can Actually Call
- Deepseek tui Tops Github Trending a Claude Code Clone Wired to Deepseeks api
Top AI Product 
Leave a comment