Dia 1.6B
A solid pick for self-hosting and full data control.
Overview
Dia 1.6B is a text-to-speech model from Nari Labs that generates dialogue directly from a transcript, with tone control, voice cloning and nonverbal sounds like laughs or sighs. Released under Apache 2.0, it remains a research-stage project: it only generates English, needs roughly 10GB of VRAM on GPU (no CPU support or quantized build yet), and voice consistency across runs isn't guaranteed without an audio prompt or fixed seed. Self-hosted, it keeps all data on the user's own infrastructure regardless of the vendor's US base.
Skill profile
Not disclosed
Strengths
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- you have GDPR constraints
Ideal use cases
- dialogue voice generation
- podcast script prototyping
- controlled voice cloning
- nonverbal audio effects
- on-prem GPU-based TTS
Access & availability
Key specifications
Privacy
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 0.9 GB | 3 GB | Any recent PC/Mac |
| Q8 · balanced | 1.7 GB | 4 GB | Any recent PC/Mac |
| FP16 · max quality | 3.2 GB | 6 GB | Any recent PC/Mac |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve nari-labs/Dia-1.6B \
--max-model-len 32768 \
--tensor-parallel-size 1 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |