Dia2 2B
Dia2 2B is the larger of Nari's two public Dia2 checkpoints. It generates English dialogue with optional speaker recordings as context. The weights use Apache 2.0, while included third-party components keep their own terms.
The larger checkpoint changes width as well as depth
Its decoder is 2,048 wide, while the 1B configuration uses 1,024. The larger model has fewer main decoder layers but a deeper auxiliary transformer. Parameter size therefore does not correspond to a simple increase in layer count.Dia2-2B model configurationDia2-1B model configuration
Reference recordings add a transcription step
The repository uses Whisper to transcribe the speaker-prefix files before generation. This adds processing time. The output includes waveform data, audio tokens and word timestamps tied to Mimi's approximately 12.5 Hz frame rate.Dia2 inference behavior and release limits
Use the same reference when comparing voices
Nari warns that voices and quality vary between generations because the model is not tuned to one voice. Its recommended path uses audio prefixes or fine-tuning for more stable output. A single successful sample does not establish consistent dialogue quality.Dia2 inference behavior and release limits
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | nari-labs/Dia2-2B |
| Published weight size | 1,919,556,608 parameters in safetensors metadata; 2B is the rounded release name.Dia2-2B weight and publication metadata |
| Repository created | 15 November 2025.Dia2-2B weight and publication metadata |
| Architecture | 28 decoder layers with a 2,048-wide embedding and a four-layer depth transformer.Dia2-2B model configuration |
| Generation limit | Up to two minutes of English speech in the documented release.Dia2-2B 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