10x Cheaper TTS at 50ms Time-to-First-Audio

Nari Labs launched a Free Public Beta of realtime text-to-speech endpoints built on its own inference engine for the Qwen3-TTS model, claiming 50 ms time to first audio in Fast mode versus 300 ms for ElevenLabs TTS v3 and a 10x cost advantage in Standard mode at $5 per 1M characters against ElevenLabs V3's $50 per 1M characters. The free endpoints, qwen3-tts-fast:free and qwen3-tts:free, each allow up to 100 requests per day with a model-agnostic concurrency limit of 2, and Nari Labs said input streaming over websocket and custom voice cloning are coming soon. Nari Labs said it built the inference engine from scratch because existing engines were not fast or cheap enough for the model, and it is offering Early Access for higher concurrency and input streaming ahead of general availability.

Since the beginning of 2024, open-weight LLMs have brought a new wind into the AI market with cheaper, yet competitive models. Now, trillions of tokens are being served using models from Z.ai, Kimi, Qwen, and DeepSeek. We believe that moment is coming to multimodal AI as well, starting with speech. Earlier this year, we realized the bottleneck was not the model, it was the inference engine. Qwen3-TTS is a phenomenal model when used correctly, but running it on existing engines is not fast and cheap enough. That’s why we built one from scratch blog post /blog/qwen3-tts-speed-cost-frontier/ , specifically tuned for the model. Our realtime TTS endpoints expand the Pareto frontier: - The world’s fastest TTS endpoint , with 50 ms time to first audio in Fast mode compared with 300 ms for ElevenLabs TTS v3.<sup 1 tts-measurement-note </sup - 10x cheaper than ElevenLabs V3 $5 per 1M characters compared to $50 per 1M characters in Standard mode. - Expressive voices, scoring above the field average in preliminary independent benchmarks. We are launching our Free Public Beta today, and everyone can use our Fast and Standard endpoints for free, only for a limited time. We think that low costs and latency for multimodality will bring so many more applications to life. If you ever wanted to add voice to your consumer application but worried about unit costs, you shouldn’t anymore. We are working hard to lower the prices even further so that everyone can get access to high quality speech. And we’re not stopping here: we have many more endpoints across speech, vision, video, and world models in the pipeline. Input streaming websocket and custom voice cloning support is coming soon, so stay tuned. Accounts have a model-agnostic concurrency limit of 2 shared across all TTS models. Daily request limits apply separately to each endpoint and reset at 00:00 UTC. Free endpoints are provided on a best-effort basis. We cannot guarantee SLOs, including uptime or latency. | Endpoint | Requests per day | |---|---| | qwen3-tts-fast:free | Up to 100 | | qwen3-tts:free | Up to 100 | We are rolling out higher concurrency and input streaming websocket support through Early Access and plan to go GA as soon as possible. If you are interested in Early Access, please contact founders@narilabs.com mailto:founders@narilabs.com . 1 tts-measurement-ref Based on preliminary independent measurements of publicly available TTS endpoints from US East. Results may vary by region, connection, and workload. Free endpoints are best effort and carry no latency or uptime SLO.