Qwen3-TTS-12Hz-0.6B GGUF
GGUF conversion of Qwen/Qwen3-TTS-12Hz-0.6B-Base — Alibaba's Qwen3-based voice-clone TTS model with a 16-codebook RVQ audio tokenizer at 12.5 Hz frame rate.
The model splits into two GGUFs:
- Backbone (
qwen3-tts-0.6b-<quant>.gguf) — the Qwen3 talker (28 layers, hidden 1024) extracted fromtalker_config. Stockqwen3arch, runs in stock llama.cpp withembeddings=true. Vocab 3072 (audio-only cb-0);tie_word_embeddings=true. - Codec + codec_lm (
codec-<quant>.gguf) — theqwen3_tts_tokenizeraudio tokenizer (12.5 Hz frame rate, 16 RVQ codebooks) bundled with theresidual_depth_arcodec_lm adaptor (4-layer Qwen3-style depth decoder + 15 codebooks_head slices + per-cb audio embed tables). Runs in codec.cpp.
This is an audio-modality codec_lm: cb-0 is sampled by the codec_lm itself (not the backbone's lm_head); the backbone's hidden state feeds the depth decoder. No text token mid-stream.
Voice cloning is driven by the speaker encoder + reference codes path. The processor pulls spk_emb (1024-dim) from a separate wav2vec2-style speaker encoder and pairs it with ref-audio codes from the codec encoder. Both must be precomputed (the speaker encoder isn't bundled here; consume Qwen/Qwen3-TTS-12Hz-Speaker-Encoder or use a precompiled speaker artefact).
Files
Backbone (qwen3-tts-0.6b-<quant>.gguf)
| File | Size |
|---|---|
qwen3-tts-0.6b-f32.gguf | 1.2 GB |
qwen3-tts-0.6b-f16.gguf | 850 MB |
qwen3-tts-0.6b-bf16.gguf | 850 MB |
qwen3-tts-0.6b-q8_0.gguf | 453 MB |
qwen3-tts-0.6b-q6_k.gguf | 351 MB |
qwen3-tts-0.6b-q5_1.gguf | 321 MB |
qwen3-tts-0.6b-q5_k_m.gguf | 302 MB |
qwen3-tts-0.6b-q5_0.gguf | 295 MB |
qwen3-tts-0.6b-q5_k_s.gguf | 295 MB |
qwen3-tts-0.6b-q4_1.gguf | 269 MB |
qwen3-tts-0.6b-q4_k_m.gguf | 257 MB |
qwen3-tts-0.6b-q4_k_s.gguf | 244 MB |
qwen3-tts-0.6b-q4_0.gguf | 242 MB |
qwen3-tts-0.6b-q3_k_l.gguf | 230 MB |
qwen3-tts-0.6b-q3_k_m.gguf | 209 MB |
qwen3-tts-0.6b-q3_k_s.gguf | 186 MB |
qwen3-tts-0.6b-q2_k.gguf | 161 MB |
Codec + codec_lm (codec-<quant>.gguf)
| File | Size |
|---|---|
codec-f32.gguf | 933 MB |
codec-f16.gguf | 608 MB |
codec-q8_0.gguf | 565 MB |
codec-q5_k_m.gguf | 548 MB |
codec-q4_k_m.gguf | 542 MB |
Inference shape (audio-modality codec_lm AR)
backbone (Qwen3, embeddings=true) hidden h
→ codec_lm_step_begin(state, h)
→ for cb in 0..15: codec_lm_step_logits → sample → codec_lm_step_push_code
→ codec_lm_step_finish → codes[16]
→ codec_lm_compose_audio_embd(codes) → next-step embedding
→ feed via b.embd; loop until EOS
Output codes (T × 16) feed directly into codec_decode to produce 24 kHz mono PCM.
Sources
- Upstream model:
Qwen/Qwen3-TTS-12Hz-0.6B-Base - Conversion tooling:
mybigday/codec.cpp(prep_qwen3_tts+lm_adaptor/qwen3_tts.py) - Inference runtime:
mybigday/llama.rn