How to Run Qwen3-TTS-12Hz-1.7B-VoiceDesign Locally: Voice Design From a Text Description

Qwen3-TTS-12Hz-1.7B-VoiceDesign is a 1.92 billion parameter open weight text-to-speech model from Qwen that builds a voice from a written description instead of a reference clip, streams audio in as little as 97ms, and covers 10 languages. Here is what it takes to run it and how to design your first voice.

7 minIntermediate

Qwen3-TTS-12Hz-1.7B-VoiceDesign is a 1.92 billion parameter open weight text-to-speech model from Qwen, released January 21, 2026. Instead of cloning a voice from a reference clip, you describe the voice you want in plain language and it generates speech to match. It covers 10 languages and streams audio with latency as low as 97ms.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

Qwen3-TTS-12Hz-1.7B-VoiceDesign entered the top 20 of Hugging Face's text-to-speech trending list. At time of writing it has 414 likes and 298,156 downloads on the Hugging Face repo, plus 13,398 stars on the QwenLM/Qwen3-TTS GitHub repo it ships from and a demo Space, Qwen/Qwen3-TTS, with 2,243 likes of its own if you want to hear it before installing anything.

Specs at a glance

Field	Value
Publisher	Qwen
Parameters	1.92 billion, per the HF API's safetensors metadata
Pipeline	Text-to-speech: written text and a voice instruction in, spoken audio out
Tokenizer	Self-developed Qwen3-TTS-Tokenizer-12Hz for acoustic compression
Languages	Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
Streaming latency	As low as 97ms to first audio
License	Apache-2.0
Release date	January 21, 2026
GitHub stars	13,398 (QwenLM/Qwen3-TTS)

License is fully open

Both the code and the weights ship under Apache-2.0, so there is no separate commercial tier or paid unlock to worry about, unlike some other open-weight TTS models that split code and weight licensing.

Hardware: what 1.92B parameters actually costs you

Qwen hasn't published a VRAM table for this checkpoint, so the honest way to size hardware is to work from the weights themselves. At 1.92 billion parameters in BF16, the weights alone take about 3.8GB (two bytes per parameter). That's the floor: loading the model needs a GPU with at least that much free VRAM, before you add anything for the codebook cache or a desktop environment running alongside it.

In practice that puts an 8GB card comfortably ahead of what this model needs for single-stream generation, since a small TTS model's runtime overhead is nowhere near what a similarly sized language model carries. If you're deciding what to build a local-AI box around, see our guide on setting up local AI on 8GB of VRAM.

What makes VoiceDesign different from voice cloning

Most open TTS models clone a voice: you feed them a reference clip and its transcript, and they copy the timbre. VoiceDesign skips the reference clip entirely. You write an instruction describing the voice, timbre, emotion, and pacing you want, and the model generates speech that matches the description rather than an existing recording. Qwen's own architecture note calls this natural language-based voice control, and it sits alongside sibling checkpoints in the same 12Hz family: CustomVoice for picking from pre-built speakers, and Base for cloning from a reference clip.

The one benchmark Qwen has published for this model family is a Seed-TTS test-set word error rate, measured on the sibling Base checkpoint rather than VoiceDesign itself: 0.77 for Chinese, 1.24 for English. VoiceDesign shares the same tokenizer and backbone, tuned instead for instruction-driven voice creation over accuracy-focused cloning, so treat that number as a family baseline rather than a VoiceDesign-specific score.

Install it

Qwen3-TTS ships as a pip package with an isolated conda environment recommended, plus an optional flash-attention build for faster inference.

zsh - install qwen-tts

$conda create -n qwen3-tts python=3.12 -y

$conda activate qwen3-tts

$pip install -U qwen-tts

optional, for faster inference:

$pip install -U flash-attn --no-build-isolation

Design your first voice

Load the VoiceDesign checkpoint, then call generate_voice_design with the text to speak and an instruction describing the voice. No reference audio, no transcript to match, just a written description.

qwen3_tts_voice_design.py

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

wavs, sr = model.generate_voice_design(
    text="Welcome aboard, we've been expecting you.",
    language="English",
    instruct="A calm, low-pitched older man speaking slowly and warmly",
)
sf.write("output.wav", wavs[0], sr)

Write the instruction like a casting note

The instruct field responds better to concrete descriptions than vague ones. Naming an age range, pitch, pace, and mood ('a calm, low-pitched older man speaking slowly') gives the model more to work with than a single adjective like 'friendly.'

Serving it behind an API

For anything beyond one-off scripts, the package ships a local web UI and server you can point other tools at instead of writing your own inference loop.

zsh - launch the demo server

$qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --ip 0.0.0.0 --port 8000

Pass --ssl-certfile and --ssl-keyfile if you need HTTPS for remote access. If you'd rather not run a GPU at all, Alibaba's DashScope also hosts Qwen3-TTS as a real-time API, covering the same checkpoints without any local install.

FAQ

What is Qwen3-TTS-12Hz-1.7B-VoiceDesign used for?

Generating speech in a voice you describe in words rather than one you record. It's built for cases where you need a specific-sounding voice (age, pitch, mood, pace) but have no reference clip to clone, across 10 supported languages.

How many parameters does it have?

1.92 billion parameters, per the safetensors metadata in the Hugging Face API. Its weights total roughly 3.8GB in BF16.

What GPU do I need to run it?

An 8GB card is enough for single-stream generation. The BF16 weights need about 3.8GB, and a small TTS model like this carries far less runtime overhead than a language model of similar parameter count.

How is VoiceDesign different from the CustomVoice or Base checkpoints?

All three share the same 12Hz tokenizer and backbone. CustomVoice picks from a library of pre-built speakers, Base clones a voice from a reference clip and its transcript, and VoiceDesign builds a voice from a written natural-language description instead.

Can I use it commercially?

Yes. Both the code and the model weights are licensed under Apache-2.0, with no separate paid tier required for commercial use.

Where to go from here

Building a local-AI box around a small model like this? How to set up local AI on 8GB of VRAM covers the hardware tier this model runs comfortably on.
Need a voice cloned from a reference clip instead of designed from scratch? Run Breeze TTS 2 locally covers cloning, and run Kokoro-82M locally covers a smaller, cheaper-to-run alternative.
Want to compare it against other open TTS models first? Browse the local AI model directory to see specs side by side before you commit disk space to a download.

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)