whatmodel.sh / tts / Dia2 2B
Dia2 2B
↗ share
+ contribute
Dia2 2B is an open-weights text-to-speech model from Nari Labs, released November 2025. Its published size is 1.92B parameters. Licence: apache-2.0.
Catalogue identity notes
Original Nari Labs 2B streaming dialogue model, distinct from Dia 1.6B and Dia2 1B. Exact model size and repository corroborate Singh et al. arXiv:2601.20510v1 section 3.1 and reference 19. Official card has a multi-step uv/CUDA setup; no tested one-command install or RAM guarantee supplied.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
HF Hub safetensors total for the official repo (pinned revision), 1,919,556,608 = model.safetensors (the name's '2B' is the rounded size). The Mimi audio codec loaded from its own repo is not included. Observed 2026-10-07.
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Dia2 2B, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 97 |  Vevo (Amphion) | 46±3 | local |
| 98 |  Qwen3-TTS 12Hz 0.6B CustomVoice | 46±6 | comparison only |
| 99 | Dia2 2B | 45±6 | local |
| 100 |  Amazon Polly Neural | 45±4 | api $16.00/M chars |
| 101 |  KittenTTS Nano 0.8 | 44±5 | comparison only |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests1 not in a task · 85% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| DailyDialog TTS · Whisper-large WER (UncovAI deepfake study) | 6.5% ✎ | ▄··········█ beats 1/2 | 85% |
$ where to run it
Other ways to run it
- Run it as published · needs ~8.07 GB, NVIDIA only · from github.com/nari-labs (official)
git clone https://github.com/nari-labs/dia2.git && cd dia2 && uv sync && uv run -m dia2.cli --hf nari-labs/Dia2-2B --input input.txt --cfg 6.0 --temperature 0.8 --cuda-graph --verbose output.wav
Official weights and documentation: huggingface.co/nari-labs ↗
$ how the score was worked out · Audio Deepfake Detection in t… 85%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Audio Deepfake Detection in the Age of Advanced Text-to-Speech models (arXiv:2601.20510v1) | 1 | 85% |
| prior · pull toward the middle | 15% | |
| weighted average | = 44.8 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json