---
license: apache-2.0
library_name: mlx
tags:
  - mlx
  - mlx-speech
  - tts
  - voice-cloning
  - audio-editing
  - apple-silicon
base_model: stepfun-ai/Step-Audio-EditX
---

# Step-Audio-EditX — MLX 4-bit

Self-contained **affine 4-bit MLX** bundle of [Step-Audio-EditX](https://huggingface.co/stepfun-ai/Step-Audio-EditX) for Apple Silicon.

This is **not** the CUDA/vLLM checkpoint [`stepfun-ai/Step-Audio-EditX-AWQ-4bit`](https://huggingface.co/stepfun-ai/Step-Audio-EditX-AWQ-4bit). Load it with [mlx-speech](https://github.com/appautomaton/mlx-speech).

Converted from [`appautomaton/step-audio-editx-8bit-mlx`](https://huggingface.co/appautomaton/step-audio-editx-8bit-mlx) using MLX affine quantization (`bits=4`, `group_size=64`, `mode=affine`).

| Component | Precision |
| --- | --- |
| Step1 LM, flow-model, VQ02 | affine int4 |
| VQ06, HiFT, CampPlus, flow conditioner | bf16 (unchanged) |

Total size ≈ **2.53 GB** (upstream int8 ≈ 4.4 GB).

## Use

```bash
pip install mlx-speech
hf download kingfang008/step-audio-editx-4bit-mlx --local-dir ./step-audio-editx-4bit-mlx

mlx-speech tts \
  --model ./step-audio-editx-4bit-mlx \
  --reference-audio reference.wav \
  --reference-text "Transcript of the reference audio." \
  --text "New cloned speech." \
  -o output.wav
```

Requires an Apple Silicon Mac (M1 or later).