Dia 1.6B
Dia 1.6B generates English dialogue from a written script. Speaker markers control turns, and an audio prompt can guide the voices. It can generate non-speech sounds such as laughter as part of the conversation.
A dialogue script controls the speaker turns
The input uses [S1] and [S2] to mark speakers. Parenthesized cues can request laughter or coughing, though the card warns that some cues can produce unexpected output. The model is not tuned to one fixed voice, so repeated generations can sound different.Original Dia-1.6B model card
Voice prompting adds a reference, not a fixed speaker identity
For voice conditioning, the original example pairs reference audio with its transcript before the new script. The runtime also downloads Descript Audio Codec. This external codec is part of the audio path; the 1.6B label describes the Dia model.Original Dia-1.6B model card
Keep the original release separate from the June revision
The official repository now points to the 0626 revision and its Transformers integration. Use the exact original identifier when reproducing work made with this earlier release.Dia release history and research credits
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | nari-labs/Dia-1.6B |
| Published weight size | 1,611,160,576 parameters in the repository's safetensors metadata. The 1.6B name is rounded.Original Dia repository metadata |
| Repository created | 20 April 2025.Original Dia repository metadata |
| Language and license | English; Apache 2.0. The model card describes research and educational use and includes restrictions against impersonation and deceptive use.Original Dia-1.6B model card |
| Published memory guidance | Around 10 GB of GPU memory for the original full version. This is the developer's example, not a measured requirement for every runtime.Original Dia-1.6B model card |
Related models
- Dia 1.6B 0626
Nari Labs · Dia
The June 2025 Dia revision with documented Transformers integration, English dialogue generation and audio-conditioned voices.
Launch date not confirmedDownloadable The smaller Dia2 checkpoint for English streaming dialogue, with reference voices and a two-minute generation limit.
Launch date not confirmedDownloadable