---
license: other
license_name: stability-ai-community-license
license_link: https://stability.ai/community-license-agreement
base_model: stabilityai/stable-audio-3-medium
tags:
  - audio
  - music
  - onnx
  - gguf
  - hot-step-cpp
---

# HOT-Step-CPP StableStep Models

Model files for the **StableStep** post-processing feature of
[HOT-Step-CPP](https://github.com/scragnog/HOT-Step-CPP) — a local AI music
generation app. StableStep re-renders the instrumental of a generated track
through **Stable Audio 3** (SDEdit-style partial re-noising) to replace
autoencoder fizz with real spectral detail, while vocals are separated,
cleaned, and remixed untouched.

These files are **converted derivatives** of
[stabilityai/stable-audio-3-medium](https://huggingface.co/stabilityai/stable-audio-3-medium)
(diffusion transformer + SAME-L autoencoder + T5Gemma text encoder), exported
to ONNX and GGUF for native C++ inference. No weights were fine-tuned; only
format conversion and precision changes (BF16/F16 storage, fp32-sensitive
tensors kept at F32) were applied.

**The easiest way to use these is HOT-Step-CPP's built-in Model Manager**
(StableStep tab), which downloads the set you need and places files
automatically.

## Backends

| Set | Files | Size | Runs on |
|---|---|---|---|
| **GGML** (recommended) | `sa3-*.gguf` (4) | ~5.8 GB | CUDA, Vulkan, CPU (Metal untested) |
| **ONNX** | `sa3-*.onnx` + `sa3-dit.onnx.data` (6) | ~10.9 GB | NVIDIA (ONNX Runtime CUDA / TensorRT) |
| **Tokenizer** (required by BOTH) | `tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json` | ~35 MB | — |

In current testing on an RTX 5090 the GGML backend is both smaller and
substantially faster than the ONNX/TensorRT path for this workload.

## Manual install layout

Relative to your HOT-Step-CPP `models/` directory:

```
models/
├── sa3-dit-BF16.gguf              ← GGML set (repo root files)
├── sa3-same-enc-F16.gguf
├── sa3-same-dec-F16.gguf
├── sa3-text-enc-BF16.gguf
└── onnx/sa3/
    ├── sa3-dit.onnx               ← ONNX set
    ├── sa3-dit.onnx.data
    ├── sa3-same_encoder.onnx
    ├── sa3-same_decoder.onnx
    ├── sa3-text_encoder.onnx
    ├── sa3-seconds_embedder.onnx
    ├── tokenizer.json             ← tokenizer files: required for BOTH backends
    ├── tokenizer_config.json
    └── special_tokens_map.json
```

## License

These weights are derived from Stable Audio 3 and are distributed under the
**[Stability AI Community License](https://stability.ai/community-license-agreement)**
(see `LICENSE-STABILITY-AI-COMMUNITY.md` / NOTICE in this repo). In short:
free for research, non-commercial, and commercial use by individuals and
organizations with less than $1M in annual revenue; larger organizations
require an [enterprise license from Stability AI](https://stability.ai/enterprise).
By downloading these files you agree to those terms. HOT-Step-CPP presents
this license for acceptance before downloading in-app.

**Powered by Stability AI.**

## Acknowledgements

- [Stability AI](https://stability.ai) — Stable Audio 3 (DiT + SAME-L autoencoder)
- [Google](https://huggingface.co/google/t5gemma-b-b-ul2) — T5Gemma text encoder (via the Stable Audio 3 release)
- Conversion + native inference: the HOT-Step-CPP project