Nari Labs

Nari Labs is a multimodal inference platform specializing in speech — their Qwen3-TTS and Qwen3-ASR models are optimized for production with latency numbers that lead independent benchmarks. The Fast TTS endpoint hits sub-50ms server-side time-to-first-audio, and the Fast STT endpoint delivers sub-40ms time-to-first-token. Both are currently in free public beta.

The pricing is aggressive: $5 per million characters for standard TTS (10x cheaper than ElevenLabs), and $0.06 per hour for standard STT (significantly under Deepgram). The Fast tier doubles the price but halves the latency. All endpoints support streaming.

Beyond APIs, Nari Labs offers dedicated multimodal inference for enterprise (model-specific runtimes with custom kernels, batching, and quantization) and fine-tuning services for multimodal models. The team behind Dia — an open-source dialogue TTS model with 2M+ downloads and 20K+ GitHub stars — and Narvatar, a human-centric avatar system under 5B parameters. Y Combinator backed, operating from San Francisco and Seoul.