Run Qwen3-TTS Locally: Free Voice Cloning in 3 Seconds

How do I install Qwen3-TTS?

Three steps. About 10 minutes plus the model download.

Step 1: Set up a clean Python env

The README is explicit about this — use Python 3.12+ in an isolated env to avoid dependency conflicts:

bash
python -m venv qwen-tts
source qwen-tts/bin/activate  # Windows: qwen-tts\Scripts\activate
pip install --upgrade pip

Step 2: Install Qwen3-TTS and FlashAttention

bash
pip install qwen3-tts
pip install flash-attn --no-build-isolation

FlashAttention 2 is recommended by the maintainers because it cuts GPU memory usage. On limited-RAM systems, set MAX_JOBS=4 before installing.

Step 3: Pick your model

The Qwen3-TTS family includes five model variants:

Model	Use case	Size
Qwen3-TTS-12Hz-1.7B-Base	General TTS	1.7B
Qwen3-TTS-12Hz-1.7B-VoiceDesign	Design a voice from scratch	1.7B
Qwen3-TTS-12Hz-1.7B-CustomVoice	Clone from a 3-second sample	1.7B
Qwen3-TTS-12Hz-0.6B-CustomVoice	Lightweight clone	0.6B
Qwen3-TTS-12Hz-0.6B-Base	Lightweight general	0.6B

For voice cloning on a laptop, start with the 0.6B CustomVoice variant.

Step 4: Clone a voice from a 3-second sample

python
from qwen3_tts import Qwen3TTS

model = Qwen3TTS.from_pretrained("Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice")
audio = model.clone(
    reference_audio="my-voice-3sec.wav",
    text="This is the voice I just cloned in three seconds.",
    language="en"
)
audio.save("output.wav")

Step 5: Stream in real time (optional)

The streaming mode advertises latency "as low as 97ms" — fast enough for live agents:

python
for chunk in model.stream(text=long_text, voice_id="my-clone"):
    play(chunk)

What is Qwen3-TTS?

Qwen3-TTS is Alibaba's open-source text-to-speech family, released January 22, 2026 under Apache 2.0. The README pitches it directly: "stable, expressive, and streaming speech generation, free-form voice design, and vivid voice cloning". It supports 10 languages out of the box: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.

Voice cloning needs just a 3-second audio reference — shorter than ElevenLabs' minimum.

Is Qwen3-TTS really free?

Yes. The license is Apache 2.0 including commercial use. You download the weights from Hugging Face, run them locally, and there is no usage cap. The repo has 11.2k stars and 1.5k forks as of May 2026 — heavy production adoption.

By contrast, ElevenLabs' Starter is $6/month for 30k credits with Instant Voice Cloning. Their Creator tier at $11/month gets you 121k credits and Professional Voice Cloning. To get high-quality 44.1kHz output you need Pro at $99/month. And every generation counts against your credit pool.

How is Qwen3-TTS different from ElevenLabs?

Three concrete differences:

Benchmarks: On the SEED-TTS test set, Qwen3-TTS-12Hz-1.7B-Base hits 0.77 WER on Chinese and 1.24 WER on English, beating CosyVoice 3, MiniMax-Speech, and FireRedTTS 2. ElevenLabs does not publish equivalent benchmarks.
Cost per generation: Qwen3-TTS is zero. ElevenLabs Pro is $99/month for 600k credits, which is roughly 10 hours of TTS — past that, overage.
Voice cloning sample: Qwen3-TTS clones from 3 seconds. ElevenLabs Instant Voice Cloning needs around 1 minute, and Professional Voice Cloning takes 30+ minutes of clean reference audio.

The maintainers' own line: "Extreme Low-Latency Streaming Generation" at sub-100ms.

What's the catch?

GPU required. You need at least 6-8GB VRAM for the 0.6B variant and roughly 12GB for the 1.7B. A laptop iGPU will not cut it.
English voice quality is excellent but not yet better than ElevenLabs Pro for character work. For audiobook narration with one specific voice, ElevenLabs Pro's Professional Voice Cloning is still slightly more natural. For everything else — agents, TTS for content, multilingual — Qwen3 wins.
Setup is one command but FlashAttention can be finicky. If pip install flash-attn fails, fall back to pip install qwen3-tts alone — it works without flash-attn, just uses more VRAM.

Bottom line

For 95% of TTS use cases — agents, content narration, multilingual voice work, voice cloning from short samples — Qwen3-TTS replaces ElevenLabs entirely. Free, runs on your laptop, 10 languages, 3-second clones. The only reason to keep ElevenLabs is character-grade audiobook narration where every syllable matters.

Related Nanoflow guides

Want a hand?

Book a 30-min call.

Walk through your stack with us. We'll find the bottleneck and map out the exact wiring you need — free.

Book the callFree intro · 30 min · cal.com