ilintar/moss-tts-gguf

MOSS-TTS-v1.5 BF16 and Q8_0 GGUF for OpenMOSS

GGUF weights for OpenMOSS-Team/MOSS-TTS-v1.5, converted for the OpenMOSS C++/GGML runtime.

Precision correction

All files in this revision are FP16-free. The full backbone follows the official BF16 checkpoint. The recommended compact backbone uses Q8_0 matrices, retains the token embedding table as BF16, and keeps normalization tensors in FP32. The shared audio sidecar stores ordinary learned weights as BF16 while retaining normalization and residual-vector-quantizer tensors in FP32.

Files

File	Tensor inventory	SHA-256
moss-tts-1.5-q8_0.gguf	253 Q8_0, 1 BF16, 145 FP32, 0 FP16	83a7885739fe92113208c2abbb45985da30ac1a40c31d4b1a2e722a8ec5256ef
moss-tts-1.5-q8_0.extras.gguf	618 BF16, 1,046 FP32, 0 FP16	d50f71cffd44e4b76db58e5be09155200b74c6058c5ebeac58e026926c9a72eb
moss-tts-1.5.gguf	254 BF16, 145 FP32, 0 FP16	f732033d4f1f528ec49ec93904a97c7c61de9ba76e3a0b63d199518dd1ea35f7
moss-tts-1.5.extras.gguf	618 BF16, 1,046 FP32, 0 FP16	d50f71cffd44e4b76db58e5be09155200b74c6058c5ebeac58e026926c9a72eb

Download one backbone and its same-stem *.extras.gguf sidecar. The Q8_0 pair is recommended for normal use; the full BF16 pair is the reference conversion.

Source revisions

• TTS: OpenMOSS-Team/MOSS-TTS-v1.5@cdd3b911b1585e3f2dbc7775ef10f9926f58850a
• Audio tokenizer: OpenMOSS-Team/MOSS-Audio-Tokenizer@3cd226ba2947efa357ef453bcad111b6eafba782
• Converter/runtime base: pwilkin/openmoss@d1f2976a89fd51bf949d28320af401e523b78c67, plus BF16-preserving conversion and BF16/FP32 codec-sidecar loading

Runtime requirement

Use an OpenMOSS build with BF16/FP32 codec-sidecar loading support. Older builds that require moss.codec.quantizer.* tensors to be FP16 cannot decode these corrected sidecars. The upstream audio tokenizer keeps its residual-vector-quantizer math in FP32, so down-converting those tensors is not a valid compatibility fix.

Validation

• Tensor audit: zero FP16 tensors in all four files.
• Full BF16 pair: CPU model load, natural end-of-speech, and codec decode to a 2.16-second WAV.
• Q8_0 pair: CPU and Vulkan model/sidecar load, natural end-of-speech, and complete codec decode.
• The long Q8_0 pilot produced 15.52 seconds on CPU and 14.48 seconds on Vulkan. CPU Whisper recovered the complete three-sentence script, including the final sentence from 10.76 seconds onward and the final words stirring once more.
• Long-pilot waveform QA found zero clipped samples, no late crop transient, and 0.13 seconds of trailing silence.

OpenMOSS usage

Plain TTS:

moss-tts-cli \
  --model moss-tts-1.5-q8_0.gguf \
  --text "Moonlight falls across the old stone road." \
  --output speech.wav

Voice cloning:

moss-tts-cli \
  --model moss-tts-1.5-q8_0.gguf \
  --reference reference.wav \
  --text "The gate is open. We should move before dawn." \
  --output cloned.wav

License

The source model and these converted weights are provided under Apache 2.0.