MOSS-SoundEffect-v2.0 BF16 GGUF for OpenMOSS
GGUF weights for OpenMOSS-Team/MOSS-SoundEffect-v2.0, converted for the OpenMOSS C++/GGML runtime.
Precision correction
Both files in this revision are FP16-free. The official pipeline runs the model with torch.bfloat16; the previous GGUF conversion stored the Qwen3 text encoder, DiT, and most VAE tensors as FP16. This rebuild keeps ordinary learned weights in BF16 without an intermediate FP16 cast. One-dimensional normalization/bias tensors and the VAE's numerically sensitive Snake parameters remain FP32.
Required files
| File | Tensor inventory | SHA-256 |
|---|---|---|
moss-soundeffect-2.0.gguf | 198 BF16, 113 FP32, 0 FP16 | 676d3ead6b12e99ecff0a3c29172f0f3571b17834724b19559c4f8572135a8de |
moss-soundeffect-2.0.extras.gguf | 376 BF16, 597 FP32, 0 FP16 | f905b2bb3be6e21908778e3d8ef8b0d06dc419cab1d3bb7613a29dc21d924b64 |
Both files must remain beside each other. The backbone is the Qwen3 text encoder; the sidecar contains the diffusion transformer, DAC VAE decoder, and scheduler metadata.
Source revision
- SoundEffect:
OpenMOSS-Team/MOSS-SoundEffect-v2.0@e35df4d82fbe87fcd5d14e5d100e349c0c3c076d - Converter/runtime base:
pwilkin/openmoss@d1f2976a89fd51bf949d28320af401e523b78c67, with BF16-preserving sidecar conversion
Validation
- Tensor audit: zero FP16 tensors in both files.
- Vulkan: complete 100-step generation and VAE decode of a three-second footstep pilot.
- CLAP semantic ranking gave the intended footstep label 0.9367 probability, ahead of the nearest alternate at 0.0445.
- Waveform QA found no clipping, late transient, or leading/trailing-silence warning.
OpenMOSS usage
moss-tts-cli \
--model moss-soundeffect-2.0.gguf \
--text "A single leather boot steps firmly onto a damp stone floor." \
--seconds 3 \
--output footstep.wav
License
The source model and these converted weights are provided under Apache 2.0.