Back to Gap Map Model components / Speech & audio

Dia

Nari Labs
open weights / Overall score: 1.4

Nari Labs' dialogue text-to-speech models, which turn a transcript tagged with two speakers into English conversation, including nonverbal sounds such as laughter and coughing. Dia2, in 1B and 2B sizes, streams: it starts speaking before the whole text arrives and can be conditioned on earlier audio, up to two minutes per generation. The first Dia, a 1.6B model, generates the dialogue in one pass.

Openness

3 high confidence
3.0
weights
open (Dia2 1B and 2B checkpoints on the Hub, ungated)
data
closed (no training-data statement in the README or on the cards)
code
partial (inference code only)
license
Apache-2.0 (Dia2 weights and code)+CC-BY-4.0 (the Kyutai Mimi codec, which Dia2 downloads on first run)
superseded-release
Dia-1.6B (2025, Apache-2.0, still published)

Dia2 downloads freely under Apache-2.0, and the Kyutai audio codec it needs is CC-BY-4.0, which asks only for attribution. Nari Labs publishes inference code only and does not describe its training data.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing month, summed across the two Dia2 and two Dia 1.6B checkpoints, most of them for the June 2025 Dia 1.6B update.

Capability

1 medium confidence
1.0

Dia has no listener rating and publishes no results on a public test set, so nothing places it beside the rated synthesizers. It sits with MeloTTS, another open voice with no published evaluation.

Verified 2026-11-12