Skip to page content
SAVRN
Search Plan a project

Open-weight model · Text to speech

step-audio-editx-8bit-mlx

by App Automaton appautomaton/step-audio-editx-8bit-mlx

step-audio-editx-8bit-mlx is an open-weight model for text to speech from App Automaton, released under Apache License 2.0. It has 3.5B parameters. At 16-bit it needs about 8.5 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 151 downloads a month.

This repository contains a self-contained pure-MLX int8 conversion of Step-Audio-EditX for local voice cloning and expressive audio editing on Apple Silicon.

Parameters3.5B
Context—
Weights4.4 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads151

Runs On

What it takes to serve step-audio-editx-8bit-mlx (3.5B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 7.1 GB 8.5 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 3.5 GB 4.2 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 1.8 GB 2.1 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.

step-audio-editx-8bit-mlx on every accelerator the SAVRN Index prices, at every precision

Model Card

By App Automaton, published under apache-2.0, revision 37b7ba38fefe.

This repository contains a self-contained pure-MLX int8 conversion of Step-Audio-EditX for local voice cloning and expressive audio editing on Apple Silicon. All pipeline components are stored as.safetensors — no PyTorch, ONNX, or NumPy files are required at inference time. This bundle is self-contained — all weights are packaged in one repository. Five-stage pipeline, all running pure MLX with bf16 activations: 1. Step1 LM (3.5B params, int8) — autoregressive dual-codebook token generation 2. CampPlus (bf16) — speaker embedding extraction from reference audio 3. Flow conditioner (bf16) — conditions generation on speaker embedding 4. Flow model (int8) — flow-matching mel spectrogram…

Read App Automaton's full model card

Step-Audio-EditX — MLX 8-bit

This repository contains a self-contained pure-MLX int8 conversion of Step-Audio-EditX for local voice cloning and expressive audio editing on Apple Silicon. All pipeline components are stored as .safetensors — no PyTorch, ONNX, or NumPy files are required at inference time.

Model Details

  • Developed by: AppAutomaton
  • Upstream model: stepfun-ai/Step-Audio-EditX
  • Task: zero-shot voice cloning, expressive audio editing
  • Runtime: MLX on Apple Silicon
  • Precision: int8 for Step1 LM, Flow model, and VQ02 tokenizer; bf16 for the rest
  • Total size: ~4.1 GB (down from ~7.7 GB upstream)

Bundle Contents

This bundle is self-contained — all weights are packaged in one repository.

File Component Format Size
model.safetensors Step1 LM (3.5B params) int8 3.5 GB
flow-model.safetensors Flow model (DiT + conformer) int8 181 MB
vq02.safetensors VQ02 audio tokenizer int8 162 MB
vq06.safetensors VQ06 audio tokenizer bf16 249 MB
hift.safetensors HiFT vocoder bf16 40 MB
campplus.safetensors CampPlus speaker embedding bf16 13 MB
flow-conditioner.safetensors Flow conditioner bf16 2.5 MB
config.json Step1 LM config + quantization JSON —
flow-model-config.json Flow model config JSON —
vq02-config.json, vq06-config.json Tokenizer configs JSON —
step-audio-tokenizer-assets.safetensors VQ02 codebook + CMVN FP32 ~2 MB
step-audio-tokenizer-config.json Tokenizer runtime config JSON —
frontend-config.json Prompt mel frontend config JSON —
hift-config.json, campplus-config.json, flow-conditioner-config.json Component configs JSON —
tokenizer.json, tokenizer_config.json Step1 tokenizer JSON —

How to Get Started

Download the bundle:

hf download appautomaton/step-audio-editx-8bit-mlx \
  --local-dir models/stepfun/step_audio_editx/mlx-int8

Voice cloning:

python scripts/generate/step_audio_editx.py \
  --prompt-audio reference.wav \
  --prompt-text "Transcript of reference audio." \
  -o cloned.wav \
  clone --target-text "New speech in the cloned voice."

Audio editing (change emotion):

python scripts/generate/step_audio_editx.py \
  --prompt-audio input.wav \
  --prompt-text "Transcript of input audio." \
  -o happy.wav \
  edit --edit-type emotion --edit-info happy

Supported Edit Types

Edit type Description --edit-info examples
emotion Change the emotion of speech happy, sad, angry, surprised
style Change speaking style whispering, broadcasting, formal
speed Change speaking speed fast, slow
denoise Remove noise from audio not used
vad Remove silences from audio not used
paralinguistic Add non-verbal sounds requires --target-text

Architecture

Five-stage pipeline, all running pure MLX with bf16 activations:

  1. Step1 LM (3.5B params, int8) — autoregressive dual-codebook token generation
  2. CampPlus (bf16) — speaker embedding extraction from reference audio
  3. Flow conditioner (bf16) — conditions generation on speaker embedding
  4. Flow model (int8) — flow-matching mel spectrogram generation
  5. HiFT vocoder (bf16) — mel spectrogram to waveform

The VQ02 and VQ06 tokenizers encode reference audio into dual codebook tokens consumed by Step1.

Links

License

Apache 2.0 — following the upstream license published with stepfun-ai/Step-Audio-EditX.

Configuration

Architecture
Step1ForCausalLM
Layers
32
Hidden size
3,072
Feed-forward size
8,192
Attention heads
48
Vocabulary size
74,752
Stored precision
bfloat16
Model type
step1

Identity and Version

Repository
appautomaton/step-audio-editx-8bit-mlx
Publisher
App Automaton
Task
Text to speech
Modality
Audio
Library
mlx
Parameters
3.5B parameters
Languages
en, zh, ja, ko
Revision
37b7ba38fefe1d9bb545d1476e1c12e3eeb7930b
First published
2026-04-11
Last updated
2026-09-27

Files and Weights

21 files, 4.4 GB in total. The weights are 8 files totalling 4.4 GB in safetensors.

Weights8 files · 4.4 GB
Configuration8 files · 3.9 KB
Tokenizer3 files · 6.4 MB
Documentation1 file · 4.5 KB
Repository1 file · 1.5 KB
Every file
FileTypeSizeSHA-256
campplus.safetensorsWeights13.9 MB 40d2618b1884
flow-conditioner.safetensorsWeights2.7 MB 108792964617
flow-model.safetensorsWeights190.3 MB 40fb429efee5
hift.safetensorsWeights41.6 MB 28a40adb4e2a
model.safetensorsWeights3.8 GB 7b7fb89f4e58
step-audio-tokenizer-assets.safetensorsWeights2.1 MB 9eb651ec816b
vq02.safetensorsWeights169.4 MB 4caa0d4ef375
vq06.safetensorsWeights261.3 MB 3b10c2501908
campplus-config.jsonConfiguration242 B —
config.jsonConfiguration650 B —
flow-conditioner-config.jsonConfiguration204 B —
flow-model-config.jsonConfiguration619 B —
frontend-config.jsonConfiguration136 B —
hift-config.jsonConfiguration881 B —
vq02-config.jsonConfiguration858 B —
vq06-config.jsonConfiguration338 B —
README.mdDocumentation4.5 KB —
.gitattributesRepository1.5 KB —
step-audio-tokenizer-config.jsonTokenizer559 B —
tokenizer.jsonTokenizer6.4 MB —
tokenizer_config.jsonTokenizer757 B —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
4.4 GB
Download from App Automaton

Released by App Automaton through its official repository on Hugging Face. Read the license.

Built From

  • Derived from stepfun-ai/Step-Audio-EditX
  • Described by arXiv:2511.03601
  • Quantized from stepfun-ai/Step-Audio-EditX

Memory Requirements

PrecisionWeights in memory
As published4.4 GB
16-bit7.1 GB
8-bit3.5 GB
4-bit1.8 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About step-audio-editx-8bit-mlx

How much GPU memory does step-audio-editx-8bit-mlx need?

About 8.5 GB at 16-bit and 2.1 GB at 4-bit: the weights (3.5B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run step-audio-editx-8bit-mlx on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use step-audio-editx-8bit-mlx commercially?

Yes. step-audio-editx-8bit-mlx is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Text to speech

svara-tts-v1

Kenpath Labs

svara-TTS is a developer-first multilingual TTS model for 19 languages (18 Indic + Indian English). Built on an Orpheus-style discrete audio token approach, it targets clarity, expressiveness, and low-latency on commodity GPUs/CPUs. It supports light-weight emotion/style control (e.g.,,,, ) and simple speaker identities (Language (Gender)), with zero-shot adaptation paths. Try it live on the Demo Space, or on Colab Deployment scripts and inference repo will be available soon. Watch our Github for updates - Place style/emotion tags at the end of the sentence: आज... सच में अच्छी खबर है — शाम को मिलते हैं! - Use punctuation to hint prosody (ellipses, commas, exclamation). - For technical or…

Open weights apache-2.0 3.3B parameters 131,072 tokens transformers
View model

Model · Text to speech

orpheus-3b-0.1-ft

Unsloth AI

03/18/2025 – We are releasing our 3B Orpheus TTS model with additional finetunes. Code is available on GitHub: CanopyAI/Orpheus-TTS Orpheus TTS is a state-of-the-art, Llama-based Speech-LLM designed for high-quality, empathetic text-to-speech generation. This model has been finetuned to deliver human-level speech synthesis, achieving exceptional clarity, expressiveness, and real-time streaming performances. Check out our Colab (link to Colab) or GitHub (link to GitHub) on how to run easy inference on our finetuned models. Do not use our models for impersonation without consent, misinformation or deception (including fake news or fraudulent calls), or any illegal or harmful activity. By…

Open weights apache-2.0 3.3B parameters 131,072 tokens transformers
View model

Model · Text to speech

orpheus-3b-0.1-ft

Canopy Labs

03/18/2025 – We are releasing our 3B Orpheus TTS model with additional finetunes. Code is available on GitHub: CanopyAI/Orpheus-TTS Orpheus TTS is a state-of-the-art, Llama-based Speech-LLM designed for high-quality, empathetic text-to-speech generation. This model has been finetuned to deliver human-level speech synthesis, achieving exceptional clarity, expressiveness, and real-time streaming performances. Check out our Colab (link to Colab) or GitHub (link to GitHub) on how to run easy inference on our finetuned models. Do not use our models for impersonation without consent, misinformation or deception (including fake news or fraudulent calls), or any illegal or harmful activity. By…

Access requested at publisher apache-2.0 3.8B parameters transformers
View model

Model · Text to speech

VibeVoice-1.5B

Microsoft

VibeVoice is a novel framework designed for generating expressive, long-form, multi-speaker conversational audio, such as podcasts, from text. It addresses significant challenges in traditional Text-to-Speech (TTS) systems, particularly in scalability, speaker consistency, and natural turn-taking. A core innovation of VibeVoice is its use of continuous speech tokenizers (Acoustic and Semantic) operating at an ultra-low frame rate of 7.5 Hz. These tokenizers efficiently preserve audio fidelity while significantly boosting computational efficiency for processing long sequences. VibeVoice employs a next-token diffusion framework, leveraging a Large Language Model (LLM) to understand textual…

Open weights mit 2.7B parameters transformers
View model

VibeVoice is a novel framework designed for generating expressive, long-form, multi-speaker conversational audio, such as podcasts, from text. It addresses significant challenges in traditional Text-to-Speech (TTS) systems, particularly in scalability, speaker consistency, and natural turn-taking. A core innovation of VibeVoice is its use of continuous speech tokenizers (Acoustic and Semantic) operating at an ultra-low frame rate of 7.5 Hz. These tokenizers efficiently preserve audio fidelity while significantly boosting computational efficiency for processing long sequences. VibeVoice employs a next-token diffusion framework, leveraging a Large Language Model (LLM) to understand textual…

Open weights mit 2.7B parameters 65,536 tokens transformers
View model

https://github.com/vibevoice-community/VibeVoice VibeVoice is a novel framework designed for generating expressive, long-form, multi-speaker conversational audio, such as podcasts, from text. It addresses significant challenges in traditional Text-to-Speech (TTS) systems, particularly in scalability, speaker consistency, and natural turn-taking. A core innovation of VibeVoice is its use of continuous speech tokenizers (Acoustic and Semantic) operating at an ultra-low frame rate of 7.5 Hz. These tokenizers efficiently preserve audio fidelity while significantly boosting computational efficiency for processing long sequences. VibeVoice employs a next-token diffusion framework, leveraging a…

Open weights mit 2.7B parameters transformers
View model