What Is ACE-Step 1.5? Versions, the XL Base Model and Vocal Extraction Explained

ACE-Step 1.5 is a free, open-source AI music model that turns a text prompt and lyrics into a complete song with vocals, in 50+ languages, and it is light enough to run on an ordinary gaming PC. It is co-led by ACE Studio and StepFun and released under the MIT license. It came out in January 2026, with its technical report dated 31 January 2026, and a larger XL series followed on 2 April 2026. It can also edit existing audio: restyle a song, regenerate one section, add a new instrument part, or extract a single part, such as the vocals, from a finished mix.

If you have seen "acestep" in a model name or a command and wondered what it is, this is the plain-English overview: who made it, what the different versions are, what the XL base model can do that the others can't, how vocal extraction works, what hardware you need, and how it compares to YuE2, the other leading open song model.

ACE-Step 1.5 at a Glance

	ACE-Step 1.5
What it is	An open music model that writes full songs (vocals plus backing) from a prompt and lyrics, and can edit existing audio
Made by	Co-led by ACE Studio and StepFun
Released	January 2026 (technical report dated 31 January 2026); XL series on 2 April 2026
How it works	An optional planning model sketches the song, then a diffusion model generates the audio
Model sizes	Audio model: 2B (base, sft, turbo) or 4B XL (xl-base, xl-sft, xl-turbo). Planning model: 0.6B, 1.7B or 4B
Song length	10 seconds to 10 minutes, up to 8 songs at once
Languages	50+ for lyrics
License	MIT for the code and weights
Official hardware	Under 4 GB of VRAM for the lightest setup, 12 GB or more for XL. Runs on NVIDIA, AMD, Intel, Apple Silicon or CPU
In Song Creator Pro	Windows, a 6 GB GPU (NVIDIA or AMD) or CPU only, no Python setup

Who Made ACE-Step 1.5?

ACE-Step is co-led by two companies: ACE Studio, which makes an AI singing voice workstation, and StepFun, an AI lab. The 1.5 technical report (arXiv 2602.00744) lists Junmin Gong, Yulin Song, Wenxiao Zhao, Sen Wang, Shengyuan Xu, Jing Guo and Xuerui Yang as authors.

A common mix-up: ACE Studio and ACE-Step are not the same thing. ACE Studio is a commercial app where you draw notes, type lyrics and have an AI singer perform them. ACE-Step is the free, open model that writes whole songs from a prompt. You can download and run ACE-Step without ever using ACE Studio's software. If you came here looking for the workstation, our ACE Studio alternatives page compares it with other options.

ACE-Step 1.5 is the second generation. The original ACE-Step, a 3.5B model, was open-sourced on 6 May 2025 under the Apache 2.0 license. Version 1.5 was rebuilt around a new two-part design and moved to the MIT license.

How ACE-Step 1.5 Works

ACE-Step 1.5 splits songwriting into two jobs:

1. Planning. A language model (built on Qwen3, in 0.6B, 1.7B and 4B sizes) reads your request and writes a song blueprint: lyrics, a caption describing the sound, and details such as duration, BPM, key and time signature. It can turn a short request ("a sad indie song about the sea") into a much more detailed brief.
2. Generating the audio. A diffusion model turns that blueprint into sound. (Diffusion models start from noise and refine it step by step; the project calls this part the "DiT", short for diffusion transformer.)

The planning step is optional. On small GPUs the official setup skips it and runs the audio model alone, which is a big part of why ACE-Step fits on modest hardware.

The developers report a full song in under 2 seconds on an A100 and under 10 seconds on an RTX 3090, and say its quality sits "between Suno v4.5 and Suno v5". Both are the team's own measurements.

The Model Versions, Including the XL Base

Each ACE-Step 1.5 audio model comes in three versions, and which version you pick matters more than its size:

Version	Generation steps	What it is for	What it can do
base	50, with a guidance setting	The original foundation model: the most varied results and the best starting point for custom training	New songs, restyling, section regeneration, plus extract, lego and complete
sft	50, with a guidance setting	Fine-tuned for higher-quality songs	New songs, restyling, section regeneration
turbo	8, no guidance setting	Trimmed down for speed while keeping quality high	New songs, restyling, section regeneration

Guidance controls how strictly the model follows your prompt, and it is the same idea as the guidance control in Song Creator Pro.

The 2B models came first. On 2 April 2026 the team added the XL series, a 4B-parameter version of the audio model in the same three versions: acestep-v15-xl-base, acestep-v15-xl-sft and acestep-v15-xl-turbo. According to the README, the XL weights take about 9 GB of VRAM, compared with about 4.7 GB for the 2B model.

The XL base model gets searched for most, and the table shows why. Its model card calls it "the foundation model supporting all tasks." For plain song generation the sft and turbo versions are the better choice (the model card rates their quality higher), but only the base models, 2B or XL, can do the three editing tasks: extract, lego and complete. If you want the larger model and also want to pull vocals out of a song, XL base is the only XL version that can do it.

Can ACE-Step Extract Vocals? Extract, Lego and Complete

Yes. The base models support three tasks that work on audio you already have. The official tutorial describes them like this:

• Extract pulls a single part out of a mixed song, for example the vocals.
• Lego adds a new part to existing ones, for example drums on top of a guitar recording.
• Complete builds a full backing track around a single part, for example guitar and drums under a vocal.

The code recognizes these part names for extract and lego: vocals, backing vocals, drums, bass, guitar, keyboard, percussion, strings, synth, brass, woodwinds and fx. That is a much longer list than traditional stem splitters, which usually stop at four or six stems.

This works differently from a separation tool like Demucs. ACE-Step's extract is generative: the model creates the requested part based on the song you give it, rather than filtering that part out of the original recording. As a result it can separate more instrument types, but the result can be less precise than a dedicated separator, which handles a fixed set of stems and stays closer to the original audio. Song Creator Pro's Extract tool offers both, so you can choose ACE-Step for the most detailed breakdown or Demucs for the cleanest vocal. Our guide on how to extract vocals and instruments from a song explains how to choose.

To run extract yourself in the official interface, load a base model (acestep-v15-base or acestep-v15-xl-base) instead of the default turbo, switch the generation mode to Extract, upload the song and choose the part you want. On a card with less than 12 GB of VRAM, the 2B base is the realistic option.

Hardware Requirements

The official repository recommends a setup for each amount of GPU memory:

VRAM	Audio model	Planning model
6 GB or less	2B turbo	None (audio model only)
6 to 8 GB	2B turbo	0.6B
8 to 16 GB	2B turbo or sft	0.6B or 1.7B
16 to 20 GB	2B sft or XL turbo	1.7B
20 to 24 GB	XL turbo or sft	1.7B (4B available)
24 GB or more	XL sft, or XL base for extract, lego and complete	4B

For XL, the model card lists 12 GB if you offload part of the model to system memory and compress it (INT8 quantization), 16 GB with offloading alone, 20 GB with no offloading, and 24 GB for full quality with the 4B planning model. The "under 4 GB" figure in the technical report refers to the lightest setup, without the planning model.

ACE-Step 1.5 is not NVIDIA-only. The repository supports NVIDIA (recommended), AMD, Intel and Apple Silicon GPUs as well as CPU, and needs Python 3.11 or 3.12. For help choosing a card, see our VRAM-first GPU buying guide.

The ACE-Step 1.5 License

ACE-Step 1.5 is released under the MIT license, one of the most permissive open-source licenses. You can use it, modify it and sell what you make with it, with no royalties to the developers. That is a real difference from YuE2, whose weights are licensed for non-commercial use, with an extra permission that lets individual creators earn money from their songs.

On training data, the Hugging Face model card lists licensed music, royalty-free and public domain music, and synthetic audio made by converting MIDI files to sound. The repository's disclaimer still asks users to check that their songs are original, disclose that AI was involved, and get permission where it is needed. For the wider question of who owns the output, see who owns AI-generated music. This is a summary, not legal advice.

ACE-Step 1.5 vs. YuE2

YuE2, released on 9 September 2026, is the other leading open song model, and Song Creator Pro includes both. They are good at different things:

	ACE-Step 1.5	YuE2
Released	January 2026 (XL in April 2026)	September 2026
WildSongBench score (SongBench average)	6.01	6.73 standard, 6.96 best of 8
Editable melody and chord score	No	Yes
Vocal languages	50+	English, Mandarin, Japanese, Korean (details)
Stem extraction and regenerating one section	Yes (base models)	No (score edits re-render the whole song)
Speed	Faster	Slower
Weights license	MIT	CC BY-NC 4.0 plus a creator permission
Official hardware	Under 4 GB to 24 GB depending on version; NVIDIA, AMD, Intel, Apple or CPU	Linux, NVIDIA GPU with 24 GB

WildSongBench is the YuE2 team's own benchmark, so read its numbers with that in mind. In short, YuE2 scores higher on song quality and is better at structure and covers that keep the original melody. ACE-Step 1.5 is faster, covers more languages, has a simpler license, runs on more kinds of hardware and offers audio tools like extraction. Our YuE2 explainer covers YuE2 in detail, and the best open-source AI music models compares both with DiffRhythm and MusicGen.

How to Run ACE-Step 1.5

From the official repository. Install the uv package manager, clone github.com/ACE-Step/ACE-Step-1.5, run uv sync, then uv run acestep to open a web interface at localhost:7860 (or uv run acestep-api for a REST API). The repository also offers portable packages for Windows and macOS. This route gives you every model version, every task and LoRA training (teaching the model a style from a few of your own songs), and it assumes you are comfortable choosing models for your VRAM and fixing a Python environment when something breaks. The team says its GitHub Pages site is its only official website, so download only from GitHub or Hugging Face.

Online. The repository points to acemusic.ai as a free hosted version if you just want to try it. As with any online tool, your prompts and audio go to someone else's servers.

Song Creator Pro. A Windows app with ACE-Step 1.5 and YuE2 built in. It runs ACE-Step 1.5 on a 6 GB GPU (NVIDIA or AMD) or CPU only, downloads and sets up the models for you, and makes them available through its Generate, Remix, Revise, Layer and Extract tools, with seed, guidance and batch controls and MP3, FLAC or WAV export. Everything runs offline. The trade-offs: it is Windows only, a GPU is strongly recommended for reasonable speed, and it does not include every research option the repository does. For example, you can load LoRAs in the app but not train new ones. It is a one-time purchase with a commercial license, and the Microsoft Store version has a free trial. To compare it with ACE Studio's own software, see ACE Studio alternatives.