Nari Labs Launches Qwen3-TTS 1.7B Endpoint with 50ms TTFA and $5 per Million Characters
Nari Labs has launched a Qwen3-TTS 1.7B endpoint powered by a custom inference engine, delivering a 50ms TTFA and a $5 per 1M characters pricing structure for v
On September 11, 2026, Nari Labs, the engineering team behind the open-source conversational speech model Dia, officially launched a dedicated inference endpoint powered by Qwen3-TTS 1.7B. Built on the company's proprietary low-latency inference engine, the new service delivers a claimed 50ms Time-to-First-Audio (TTFA) and introduces a competitive pricing tier of $5 per one million characters, directly targeting the real-time voice agent and conversational AI developer ecosystem.
Image source: Toby Kim (@doyeob) / Nari Labs
Having earned over two million downloads and nearly 20,000 GitHub stars with their dialogue model Dia, Nari Labs is now expanding beyond model weights into commercial-grade inference infrastructure, providing developers with scalable and cost-effective voice capabilities.
Proprietary Inference Engine Integrated with Qwen3-TTS 1.7B
The technical core of the new endpoint combines the open-weights Qwen3-TTS 1.7B foundation model with Nari Labs's custom inference acceleration layer.
While modern open-source speech models have achieved expressive prosody and natural intonation, deploying them for production voice agents has traditionally presented steep engineering challenges due to GPU hardware costs and pipeline latency. Nari Labs developed a specialized runtime designed to mitigate these throughput and latency bottlenecks.
- Preserved Model Fidelity: The endpoint retains the full nuances, emotional expressiveness, and prosodic control of the 1.7-billion parameter Qwen3-TTS architecture without degrading output fidelity.
- Optimized Audio Streaming: By streaming generated audio chunks to client applications as soon as synthesis begins, the infrastructure avoids conversational pause gaps and maintains natural dialog flow.
Nari Labs stated that open-source architectures are positioned to lead not only in frontier language models but across multimodal AI disciplines, with speech serving as the initial frontier for their managed infrastructure offerings.
Latency Benchmarks, Cost Efficiency, and Production Considerations
According to benchmark numbers announced by Nari Labs, the endpoint reaches a Time-to-First-Audio metric of 50 milliseconds.
The team positions this response time as five times faster than Cartesia and roughly ten times cheaper than established commercial providers such as ElevenLabs. Because these figures represent internal benchmarks reported by Nari Labs, independent engineering evaluations under diverse client network conditions remain essential to assess consistent real-world performance.
- Pricing Structure and Availability: The service is currently accessible through an initial free trial promotion on narilabs.com, after which usage transitions to the standard rate of $5 per million characters.
- Developer Impact: For latency-critical applications such as interactive customer support agents, real-time simultaneous translation systems, and responsive gaming NPCs, the endpoint provides an accessible alternative to high-cost proprietary voice platforms.
Sources
- Toby Kim (@doyeob) Official Launch Announcement on X: Nari Labs Qwen3-TTS 1.7B Endpoint Launch Post (Official release thread detailing the 50ms TTFA benchmark, 1.7B parameter architecture, and comparative latency metrics).
- Nari Labs Official Platform and Interactive Playground: Nari Labs Official Website (Live interactive web console, product documentation, developer onboarding, and commercial API endpoint access).