Nari Labs
Nari Labs develops speech and avatar models. Its Dia releases generate spoken dialogue from scripts, while its current commercial business also optimizes models made by other developers. The company works from San Francisco and Seoul.
Company & leadership
- Legal operator
- Nari Labs, Inc., a United States company.
- Locations
- The company describes its team as working from San Francisco and Seoul. Its privacy policy lists Newark, Delaware as its headquarters and a separate San Francisco mailing address.
- Original Dia team
- Toby Kim's release account names Jaeyong Sung as his collaborator. The reviewed company pages do not state a current CEO title or an incorporation date.
- Public model history
- The original Dia-1.6B repository was created on 20 April 2025. This is a model publication record, not a company incorporation date.
- Backing
- Nari identifies Y Combinator as a backer. Its About page does not disclose an ownership percentage or a funding amount.
- Name
- The Dia repository explains that Nari is a Korean word for lily.
Dia began with control over a conversation
In his release account, Toby Kim described building Dia with Jaeyong Sung over three months. The project released weights and inference code so other developers could work with dialogue generation. That release is the clearest public starting point for the model family; it does not establish the company's legal founding date.
The original repository credits Google's TPU Research Cloud for training compute and Hugging Face for a ZeroGPU grant. It also names SoundStorm, Parakeet and Descript Audio Codec as influences. Those acknowledgements describe research inputs and infrastructure support, rather than ownership by those organizations.
Dia2 changes when speech generation can start
Dia2 can start generating audio before the complete input text arrives. Its two released checkpoints target English dialogue of up to two minutes. Nari's repository distinguishes this model behavior from the separate TTS server and Sori speech-to-speech engine, which it still lists as upcoming.
Nari's model research and its hosted APIs have different origins
Nari made its Qwen3-TTS and Qwen3-ASR APIs generally available on 17 September 2026. Standard and Fast are serving options. The open TTS implementation optimizes Qwen3-TTS; it is not evidence of a new Nari-trained foundation model. These hosted Qwen services therefore do not add model entries to the Dia catalog.
Nari also names Narvatar as its own avatar model. Its public training page provides no download or model-specific API contract.