---
license: other
library_name: ggml
tags:
- gguf
- text-to-speech
- tts
- moss-tts
- vulkan
- directml
---

# MOSS-TTS-PNY GGUF

Portable GGUF/ONNX artifacts for the MOSS-TTS-PNY C++ runtime.

This repository is intended for use with the model-free Windows Vulkan + DirectML
bundle produced by the companion C++ project. Put these files in the bundle's
`models/` directory, or run `download-models.bat` from the bundle root.

## Files

| File | Purpose |
| --- | --- |
| `moss-tts-pny-f16.gguf` | Full precision main MOSS TTS GGUF. |
| `moss-tts-pny-global-q8_0.gguf` | Main model with the global transformer quantized to Q8_0. |
| `moss-tts-pny-global-q6_k.gguf` | Main model with the global transformer quantized to Q6_K. |
| `moss-audio-decoder4-f16.gguf` | Full precision decoder4 codes-to-features model. |
| `moss-audio-decoder4-q8_0.gguf` | Q8_0 decoder4 codes-to-features model. |
| `moss-tts-qwen2-tokenizer.gguf` | Qwen2 tokenizer vocabulary GGUF used by the runtime. |
| `istftnet2_decoder.onnx` | ONNX iSTFTNet2 vocoder used by the Windows DirectML path. |
| `istftnet2-vocoder-f16.gguf` | Experimental GGUF vocoder artifact. The packaged demos currently use ONNX. |

## Windows Bundle Usage

From the extracted Windows bundle root:

```bat
download-models.bat
run-vulkan-directml-full.bat
```

The quantized demo uses:

```bat
run-vulkan-directml-quant.bat
```

Expected model layout:

```text
models/
  moss-tts-pny-f16.gguf
  moss-tts-pny-global-q8_0.gguf
  moss-tts-pny-global-q6_k.gguf
  moss-audio-decoder4-f16.gguf
  moss-audio-decoder4-q8_0.gguf
  moss-tts-qwen2-tokenizer.gguf
  istftnet2_decoder.onnx
  istftnet2-vocoder-f16.gguf
```

## Notes

- The model-free Windows zip intentionally excludes these large artifacts.
- The full demo uses `moss-tts-pny-f16.gguf` and `moss-audio-decoder4-f16.gguf`.
- The quant demo uses `moss-tts-pny-global-q6_k.gguf` and
  `moss-audio-decoder4-q8_0.gguf`.
- The ONNX vocoder is the supported vocoder path for the DirectML Windows demo.