ACE-Step 1.5 XL Turbo

ACE-Step 1.5 XL Turbo

Generates music and other audio in just 8 steps with a 4B-parameter diffusion model, run locally through a Gradio interface. For creators and developers with a GPU of 12 GB or more.

Made by ACE-Step on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ACE-Step 1.5 XL Turbo is the fast, distilled version of the 4B-parameter XL model in the ACE-Step 1.5 music generation family. It makes audio in 8 inference steps with no classifier-free guidance. ACE Studio and StepFun co-lead the project, and the weights are released under the MIT license.

Who it suits: It suits creators and developers who want to generate music on their own GPU instead of calling a hosted service.

Industry | Who uses it | How they use it | Example
Music production | Independent musicians | Generate draft tracks locally to explore ideas | Try several short ideas before recording a real version
Game development | Small studios | Produce background music with their own hardware | Create placeholder tracks for a prototype level
Video and media | Content makers | Make music beds for clips | Generate a backing track for a short video
Research | Audio AI researchers | Study a distilled diffusion transformer for music | Compare the 8-step turbo model with the 50-step ones

You can skip it if you have under 12 GB of GPU memory or want the extract, lego and complete tasks, which the card lists only for the base model.

Where to use it:

• A browser, through the Gradio interface that the run command starts
• Windows, macOS and Linux with Python
• A GPU with 12 GB or more of memory, with 20 GB recommended

Use case | Who | What they get back
Music drafts | Musicians | Generated audio produced in 8 steps
Background music | Game and video makers | Audio tracks made on your own machine
Model comparison | Researchers | Output from the turbo model to set against base and SFT

How to set it up: You need these first:

• Python and git
• A GPU with at least 12 GB of memory, using CPU offload and INT8 quantization
• Disk space for about 18.8 GB of bf16 weights

1. Get the ACE-Step 1.5 code from its GitHub project and open its folder.
2. Install it in editable mode with pip.
3. Download this model into the checkpoints folder.
4. Start the Gradio interface with this model's config.

cd ACE-Step-1.5
pip install -e .
huggingface-cli download ACE-Step/acestep-v15-xl-turbo --local-dir ./checkpoints/acestep-v15-xl-turbo
python acestep --config-path acestep-v15-xl-turbo

How to use it: You start the Gradio interface with the last command above, then use it in your browser to generate audio. The card does not describe the interface's fields in the text available here.

Tool | What it does
acestep-v15-xl-turbo | This repository: 8 steps, no guidance, standard tasks
acestep-v15-xl-base | 50 steps with guidance, and all tasks including extract, lego and complete
acestep-v15-xl-sft | 50 steps with guidance, top quality rating, medium diversity, standard tasks
acestep-5Hz-lm-0.6B | Smaller language model with medium audio understanding and composition
acestep-5Hz-lm-1.7B | Mid-size language model, included in the main release
acestep-5Hz-lm-4B | Largest language model, with strong audio understanding and composition

Example: You start the interface with the turbo config and generate a short piece of music. The model finishes in 8 steps and gives you an audio file. The card does not show sample prompts, so try the interface to learn its inputs.

For newcomers and experts: ACE-Step makes music from your instructions, on your own computer. A good first try is following the four setup steps on a GPU with at least 12 GB, then generating one short piece in the browser.

For readers who know the field, the decoder is a diffusion transformer with 32 layers, 2560 hidden size and 32 attention heads, plus an 8-layer encoder. Distillation removes the need for guidance, so 8 steps replace 50, at some cost in diversity. The 0.6B, 1.7B and 4B language models all work with the XL decoder.

Install ACE-Step and the checkpoint
Start the Gradio interface
Give your music instructions
Run 8 distilled diffusion steps
Get the generated audio

Good to know:

• GPU memory: 12 GB or more with CPU offload and INT8, 16 GB with offload, 20 GB without offload, and 24 GB for full quality with the 4B language model.
• Counts: Hugging Face lists 5B parameters and 19.9 GB in F32, while the card says about 4B and 18.8 GB in bf16.
• Training data: the card says it used licensed music, royalty-free or public domain audio and synthetic MIDI-to-audio data.
• Commercial use: the card says generated music can be used commercially, so confirm that against your own needs.

Title: Inference steps per model
acestep-v15-xl-turbo: 8 steps
acestep-v15-xl-base: 50 steps
acestep-v15-xl-sft: 50 steps

• License: MIT.
• Cost to run: none named. You need a GPU with at least 12 GB.
• Data it reaches: the downloads from GitHub and Hugging Face, plus what you put into the interface.
• Not for: the extract, lego and complete tasks, which the card lists only for the base model.

On Hugging Face, 13 November 2026

• Task: text to audio
• Library: transformers
• Parameters: 5B
• Weights: about 19.9 GB in F32; about 2.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 9,152
• Likes: 235
• License: mit
• Quantized versions: 4
• Fine-tunes: 10
• Tags: transformers, safetensors, acestep, feature-extraction, audio, music, text2music, text-to-audio

Links

• Model: https://huggingface.co/ACE-Step/acestep-v15-xl-turbo
• Files: https://huggingface.co/ACE-Step/acestep-v15-xl-turbo/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ACE-Step/acestep-v15-xl-turbo

Then follow the usage in its model card. Nothing in this kit runs by itself.