Nari Labs

Nari Labs is a Y Combinator-backed inference company building speech models and serving them behind an API tuned for latency rather than raw throughput. The team is known in the open-source community for Dia, a dialogue text-to-speech model with more than 2 million downloads, and for Narvatar, an avatar system; their public repositories have gathered over 20,000 GitHub stars.

The commercial products are two 1.7-billion-parameter models with a standard and a fast tier each:

Nari Qwen3-TTS 1.7B — text-to-speech at $5 per million characters, or $10 per million on the fast tier, which the company measures at under 50 ms server-side time-to-first-audio.

Nari Qwen3-ASR 1.7B — speech recognition at $0.06 per hour of audio, or $0.12 per hour on the fast tier, with under 40 ms server-side time-to-first-token.

Why it is interesting

The positioning is explicitly on the cost-latency frontier rather than on model size. Nari Labs reports placing on the Pareto frontier of the Coval voice benchmarks, claims first place on latency for both speech synthesis and recognition, and prices its text-to-speech at roughly a tenth of ElevenLabs. On density it reports running 80 concurrent live PersonaPlex-7B calls on a single NVIDIA H100 — the figure that matters if you are costing out a voice agent that must hold thousands of simultaneous conversations.

Beyond the shared API

Two services sit alongside the public endpoints. Dedicated inference provisions and tunes capacity for enterprise workloads that cannot share a queue, and Training fine-tunes the multimodal models on customer audio — relevant for domain vocabulary, accents and languages where general models degrade.

Nari Labs is aimed at developers building voice agents, telephony systems, dictation products and real-time assistants, where per-minute cost and the delay before the first syllable decide whether the conversation feels natural. Teams that want to self-host can start from the open models instead.