You can deploy this model and start earning money today!

notmax123/
QwenTTS-he-1.7B-GGUF

Name: QwenTTS-he-1.7B-GGUF
Availability: InStock

Qwen/Qwen3-TTS-12Hz-1.7B-Base

quantized

apache-2.0

Views

No views yet

QwenTTS-he-1.7B-GGUF

GGUF build of notmax123/QwenTTS-he-1.7B, the Hebrew LoRA for Qwen3-TTS, merged into its base and converted for qwentts.cpp. This is the form MamboTTS downloads: it runs on GGML with no Python, PyTorch or PEFT at inference time.

Files

File	Size	Role
`qwen-talker-1.7b-base-Q4_K_M.gguf`	1.22 GB	Talker LM — text to 12.5 Hz codec frames. Carries the merged Hebrew weights.
`qwen-tokenizer-12hz-Q4_K_M.gguf`	255 MB	12 Hz audio codec — frames to 24 kHz mono audio. Unmodified upstream.

Both are required.

Input is stressed IPA, not Hebrew script

The adapter was trained on phonetic input, so the talker expects stressed IPA and will not read Hebrew orthography:

ʃalˈom, mˈa ʃlomχˈa hajˈom?

Use a Hebrew G2P front end such as RenikudPlus to produce it. MamboTTS ships RenikudPlus alongside these files for exactly this reason.

Usage

console

printf 'ʃalˈom, mˈa ʃlomχˈa hajˈom?' | ./build/qwen-tts \
  --model qwen-talker-1.7b-base-Q4_K_M.gguf \
  --codec qwen-tokenizer-12hz-Q4_K_M.gguf \
  --ref-wav reference_voice.wav \
  --lang auto --greedy -o out.wav

These are Base-mode checkpoints: there are no named speakers. The voice comes from --ref-wav, a reference recording of the voice to clone.

How it was built

The upstream adapter is a LoRA that cannot be loaded by a GGML runtime, so the weights were merged first. The merge is not a plain LoRA fold — besides the 231 lora_A/lora_B pairs at alpha / r = 2.0, the adapter's modules_to_save lists 20 fully retrained tensors (codec_head, text_projection, and the fifteen code_predictor.lm_head.* heads) that replace their base counterparts outright. Folding only the LoRA pairs yields a model that loads and sounds wrong.

Merge — scripts/merge_qwen_lora.py in the MamboTTS repo, math in float32 and cast back to bf16.
Convert — convert.py from qwentts.cpp, producing F32 GGUFs.
Quantize — quantize from the same project, to Q4_K_M. RVQ codebooks and their projections stay F32 under that tool's policy, since nearest-neighbour lookup is sensitive to per-row quantization noise.

Language coverage

Hebrew only. The base model speaks ten languages, but this adapter retrained the output heads for Hebrew phonotactics, so the others are not claimed here.

License

Apache-2.0, matching both the base model and the adapter.

notmax123/QwenTTS-he-1.7B-GGUF