Qwen3-TTS 0.6B vs 1.7B
On a GPU they run at the same speed, about 1.9 seconds per second of speech, but 1.7B needs twice the memory: 4.7 GB of VRAM against 2.4 GB. On the CPU, 0.6B is faster (4.69 s against 5.57 s) and needs half the RAM. Both clone from a clip and its transcript, and both are Apache 2.0.
Measured side by side
| Qwen3-TTS 0.6B | Qwen3-TTS 1.7B | |
|---|---|---|
| VRAM at peak | 2.4 GB | 4.7 GB |
| RAM on the CPU | 5.6 GB | 10.9 GB |
| Speed on the GPU | 1.95 s per second of speech | 1.92 s per second of speech |
| Speed on the CPU | 4.69 s per second of speech | 5.57 s per second of speech |
| Clones from | clip and transcript | clip and transcript |
| License | Apache 2.0 | Apache 2.0 |
In practice: a 10-second line takes about 20 seconds on the GPU with either size, about 47 seconds on the CPU with 0.6B, and about 56 seconds with 1.7B.
Which to choose
- Start with 0.6B. It fits almost any recent graphics card and runs well on the CPU.
- Try 1.7B if you have the memory. On a GPU it costs no extra time, only VRAM.
- Judge by ear, not by size. We haven't measured quality, and it depends on the voice. Generate the same line with both using Compare engines, press B to hide which is which, and keep the better take.
Getting Qwen3-TTS in Mimic
- Open Settings › Engines and download Qwen3-TTS 0.6B, 1.7B, or both. Both sizes share one Python environment, separate from every other engine's; only the model files differ.
- Download speech-to-text too, so the transcript Qwen3-TTS needs is written for you.
- Only one engine is in memory at a time: switching unloads the other first.
Full steps are in how to clone a voice offline on Windows.