Qwen3-TTS 0.6B (12Hz) — GGUF for qwen3-tts.cpp

GGML/GGUF conversions of Qwen's Qwen3-TTS 12Hz 0.6B Base model and its 12Hz tokenizer, ready for qwen3-tts.cpp (and anything speaking its C ABI, e.g. the qwen3-local engine of rust-tts-wrapper ≥ 1.6.3).

No conversion needed: these are the exact filenames the qwen3tts C API looks for in its model directory.

Files

Use either predictor GGUF (the C API accepts f16 or q8_0 names automatically; both in the dir = f16 wins). The tokenizer is always required. Note q8_0 here is ~27% smaller rather than the theoretical 50%: the converter deliberately keeps embeddings, norms, biases and LM heads in f16 for quality (mixed quantisation per tensor role). Quality difference is not formally A/B'd yet — f16 is the reference.

10 languages: en, zh, ja, ko, de, fr, ru, pt, es, it. Zero-shot voice cloning from a short reference clip.

Usage with rust-tts-wrapper (qwen3-local engine)

Download both files into one directory, then:

{ "modelsDir": "C:/path/to/this/dir", "threads": 4 }

(the qwen3-local engine's credentials JSON; threads optional)

Note on tokenizer filenames

The current convert_tokenizer_to_gguf.py in qwen3-tts.cpp emits qwen3-tts-tokenizer-f16-v2.gguf; the C API expects the name above — rename after converting, or use these files as-is.

Provenance & licence

• Source: Qwen/Qwen3-TTS-12Hz-0.6B-Base and Qwen/Qwen3-TTS-Tokenizer-12Hz (Apache-2.0)
• Conversion: scripts/setup_pipeline_models.py / convert_tts_to_gguf.py / convert_tokenizer_to_gguf.py from qwen3-tts.cpp (willwade's fork: AACTools/qwen3-tts.cpp)
• Converted 2026-10-10 (f16) / 2026-10-23 (q8_0, same script with --type q8_0)