ACE-Step 1.5 is an open-source music model co-led by ACE Studio and StepFun, released under the MIT license. It makes full songs, with vocals or instrumental, from a text prompt and optional lyrics, from 10 seconds to 10 minutes long, in more than 50 languages. You can run it free on your own computer, or use it in a browser on RaoMusic, where it needs no paid subscription and a generation starts at 10 credits.
Last checked: November 4, 2026.
What can ACE-Step 1.5 do?
According to the official documentation and GitHub repository, the full release covers:
- Text to music: a style prompt plus lyrics with section labels, from 10 seconds to 10 minutes.
- Editing: cover versions of existing audio, regenerating a selected part of a track (repainting), and turning vocals into background music.
- Your own style: LoRA training from a handful of songs. The project quotes 8 songs and about an hour on an RTX 3090 with 12 GB of VRAM.
- Speed: the project says a full song takes under 2 seconds on an A100 and under 10 seconds on an RTX 3090.
How do you run ACE-Step 1.5 on your own computer?
Follow the official installation guide: get the code, set up Python 3.11 or 3.12, download the models, then launch the web interface. The project says the base model runs on less than 4 GB of VRAM. With more memory, its GPU table recommends adding a larger language model for better results.
Running it yourself costs nothing beyond the hardware and gives you every feature above. If you don't have a suitable GPU, or you just want to try a few prompts first, the browser version below needs no setup.
How do you use ACE-Step 1.5 online?
On RaoMusic, ACE-Step 1.5 is one of the engines in the model menu on the Create page.
- Pick ACE-Step 1.5 from the model menu.
- Choose a mode. In Simple, describe the use, mood and instruments in a sentence, and turn on Instrumental if you want a background bed. In Custom, paste your lyrics with labels like
[Verse]and[Chorus]and add a style line. Our guide to turning your own lyrics into a song shows how to format them. - Set the length anywhere from 10 to 600 seconds. The default is 120.
- Decide whether to keep Thinking on (see below), then click Generate.
Each run gives you one track: an MP3 for playback and a lossless FLAC download. To compare ACE-Step with the other engines on the same lyrics, see our AI music model comparison.
Is Thinking worth turning on?
Thinking is on by default. It is slower and costs twice as much per second, and RaoMusic describes it as usually sounding better: 0.16 credits per second with it on, 0.08 with it off, rounded up, with a 10-credit minimum. That minimum means anything up to about a minute costs 10 credits either way, so on short clips Thinking is free. At the default 120 seconds it's 20 credits instead of 10.
The bigger cost is waiting. In our test below, the Thinking run took roughly nine times as long as the run without it. Use Thinking off to sketch ideas quickly, then turn it on for the version you plan to keep.
Which advanced settings should you change?
All four are optional. Leave them on Auto for a first draft, then add one at a time when the track has to match something:
- BPM: a whole number from 30 to 300, for music that has to sit on a video cut or an existing beat.
- Key: such as C Major, Am or F# minor, when the track has to play alongside another part.
- Time signature: 2/4, 3/4, 4/4 or 6/8, for example 3/4 for a waltz.
- Vocal language: pick the language of your lyrics, or leave Auto to let the model detect it.
How did Thinking on and off compare in our test?
We ran the same Simple-mode prompt twice with Instrumental on and a 60-second length, once with Thinking and once without, and timed each run from the request until the track was ready.
Prompt: A short instrumental lo-fi hip-hop bed for a late-night coding session, warm electric piano, soft brushed drums, rounded bass, no vocals.
| Setting | Thinking on | Thinking off |
|---|---|---|
| Length | 60 seconds | 60 seconds |
| Credits | 10 | 10 |
| Time until ready | About 4 minutes (243 s) | About 28 seconds |
Thinking on
Listening note: Warm electric piano, mellow bass and a relaxed brushed beat from the start. It works as is under a coding stream or a tutorial. A faint breathy, voice-like sound appears in the last moment, so trim the final second before you publish.
Thinking off
Listening note: The same piano and bass, no vocals, but the drums only come in after about 20 seconds and sound like filtered hip-hop hits rather than brushes. It's fine for a quiet background. If you need the beat from the first bar, generate again or switch Thinking on.
Thinking on followed the prompt more closely; Thinking off was much faster and still usable. Since Thinking costs nothing extra at this length, it is the better default for short clips when you can wait a few minutes. For more prompts in this style, see our lo-fi music guide.
Frequently Asked Questions
The project is released under the MIT license, which allows commercial use of the code and model. Its README also asks users to check that their tracks are original and to disclose AI involvement. Tracks you generate on RaoMusic are covered by RaoMusic's royalty-free commercial license.
The project lists more than 50. On RaoMusic, the vocal language menu offers English, Chinese, Japanese, Korean, Spanish, French and German, plus Auto, which lets the model detect the language from your lyrics.
Not with ACE-Step itself. The browser version is text-to-music only. Cover generation, repainting and reference audio are part of the local release.
Yes. Once the MP3 copy is ready, you can extend the track, make a cover version, or add vocals or an instrumental with RaoMusic's other tools. Replacing a single section isn't available for ACE-Step tracks.
A larger variant released in April 2026 with a 4-billion-parameter decoder for higher audio quality. It is for local use and needs at least 12 GB of VRAM with offloading, or 20 GB without.