Entity
Nari Labs
Nari Labs is an open-source speech AI startup that released the Qwen3-TTS and Qwen3-ASR models along with a specialized multimodal inference engine, aiming to challenge closed-source dominance in voice AI.
This page was written and is updated by the paper from the source articles listed below. Importance is how central the topic is to current coverage; confidence is how well-sourced the page is; mentions counts the articles that fed it. How the wiki works →
Nari Labs is an open-source speech AI startup that has released Qwen3-TTS and Qwen3-ASR models, claiming high accuracy, low latency, and low cost compared to existing solutions (Hacker News/Nari Labs blog, September 14, 2026). The company previously built Dia, described as the first open-source text-to-speech model capable of natural dialogue generation (Nari Labs blog, September 14, 2026).
## Background
According to the company's own announcement, Nari Labs built a specialized inference engine for multimodal speech models because existing systems such as vLLM and SGLang are not well suited for multimodal inference (Nari Labs blog, September 14, 2026). The company has open-sourced this inference engine on GitHub. Nari Labs frames its mission as challenging the dominance of closed-source speech models by proving that the market gap is fundamentally an inference infrastructure problem rather than a model quality problem.
## Strategic Significance
The open-source speech AI space has become increasingly competitive as voice interfaces gain adoption in enterprise and consumer applications. Nari Labs' focus on inference efficiency for multimodal models — rather than just model architecture — addresses a real technical gap, as general-purpose LLM inference engines are often poorly optimized for audio generation and transcription workloads. This positions the company within a broader wave of specialized AI infrastructure startups competing against both closed-source incumbents (OpenAI, ElevenLabs) and general-purpose inference platforms.
## Key Facts
- Released Qwen3-TTS and Qwen3-ASR models emphasizing high accuracy, low latency, and low cost (Nari Labs blog, September 14, 2026)
- Previously developed Dia, an early open-source natural-dialogue TTS model (Nari Labs blog, September 14, 2026)
- Built and open-sourced a specialized inference engine for multimodal speech models, arguing that vLLM and SGLang are inadequate for this use case (Nari Labs blog, September 14, 2026)
- Leads Coval voice AI benchmarks according to company claims (Nari Labs blog, September 14, 2026)