Alibaba's Qwen3-TTS Runs on Any Hardware With This Community GGUF Port
A community GGUF port lets the Qwen3-TTS voice cloning and voice design models run on CPU, CUDA, Metal, and Vulkan through qwentts.cpp.
Takeaways
- Community GGUF port of Alibaba's Qwen3-TTS runs via qwentts.cpp on CPU, CUDA, Metal, Vulkan.
- Ships 0.6B and 1.7B talkers in F32, BF16, Q8_0, and Q4_K_M; Q4_K_M is under 1 GB.
- Three modes: named-speaker base, zero-shot voice cloning, and voice design from text descriptions.
- Covers 10 languages plus Mandarin dialects, 24 kHz mono output, 97 ms end-to-end latency upstream.
- Apache 2.0 licensed for both models and the 12 Hz audio tokenizer, commercial use allowed.
- Quick start via qwentts.cpp repo with prebuilt example scripts for each mode.
Qwen3-TTS just got a lot easier to run outside a Python/CUDA stack. A community-built GGUF port packages Alibaba's open-source speech models for qwentts.cpp, a C++17/GGML runtime that runs on CPU, CUDA, Metal, and Vulkan. The Serveurperso/Qwen3-TTS-GGUF repo has already crossed 860k downloads, reflecting strong demand for a llama.cpp-style path to a modern TTS model.
What shipped
The upstream models come from the Qwen team at Alibaba Cloud. Qwen3-TTS covers 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian) plus multiple dialectal voice profiles, with contextual understanding that lets it adapt tone, speaking rate, and emotional expression from instructions and text semantics. The GGUF release converts the 0.6B and 1.7B talker checkpoints plus the 12 Hz audio tokenizer into quantized files ranging from F32 down to Q4_K_M, with Q8_0 recommended as the default.
Two GGUFs load together: a talker (Qwen3 LM plus a code predictor head and optional speaker encoder that turns text into 12 Hz codes) and the shared tokenizer (SEANet plus ConvNeXt plus DAC v2 plus RVQ) that maps those codes back to 24 kHz mono audio. Three modes are exposed:
- base - zero-shot TTS with named speakers and Mandarin dialects (0.6B and 1.7B)