What is ACE-Step 1.5?
ACE-Step 1.5 is an open music model co-led by ACE Studio and StepFun. A small language model (0.6B, 1.7B or 4B, built on Qwen3) plans the song, writing metadata, lyrics and captions, and a diffusion transformer renders the audio. It generates songs with vocals from 10 seconds to 10 minutes long, with lyrics in more than 50 languages.
Weights and code are MIT-licensed, and the team says the training data is licensed, royalty-free or synthetic, so outputs can be used commercially. The default turbo model needs under 4 GB of VRAM. An XL series with a 4B-parameter decoder followed on 2 April 2026 for higher quality, needing at least 12 GB with offloading.
Beyond text to music it does covers, repainting of a chosen section, stem separation, adding layers, accompaniment for a sung vocal, and LoRA fine-tuning from a few songs.
Is ACE-Step 1.5 worth it?
ACE-Step 1.5 is the open music model most people can actually run and use commercially. It is MIT-licensed, the turbo model needs less than 4 GB of VRAM, and according to the README it renders a full song in under 10 seconds on an RTX 3090 and under 2 seconds on an A100. It also runs on Apple Silicon, AMD, Intel GPUs and CPU, where MiniMax Music 3 and YuE2 need NVIDIA.
It has the broadest editing toolkit of the open song models we cover. Besides generating up to 10 minutes from a prompt and lyrics, it repaints a chosen section, separates stems, adds layers to an existing track, writes accompaniment for a sung vocal and reads BPM and key from audio. Training a LoRA on 8 of your own songs takes about an hour on a 12 GB RTX 3090, the README says. The Gradio app and a REST server ship in the same repository, which has 14,478 GitHub stars.
The catch: quality sits a step below the newest open rivals. The team places it between Suno v4.5 and Suno v5; on the YuE2 team's WildSongBench it averaged 6.0118 on SongBench, against 6.2830 for MiniMax Music 3 and 6.7316 for YuE2, though its lyrics were clear at a 7.46% phoneme error rate. Its covers do not keep the original tune: it identified the source song 2.4% of the time on SHS100K, against 71.3% for YuE2 given a score. The XL decoder sounds better but needs 12 GB with offloading and 20 GB without. The licensed-data claim comes from the team, and the card names only categories of data, not sources.
Our take: for a game studio, a video maker or an app developer who needs songs they can ship from a model that fits on an ordinary GPU or a Mac, ACE-Step 1.5 is the default open choice; MiniMax Music 3 is worth the bigger GPU when vocal polish matters most.
Where it earns its place
An indie game developer with a 6 GB laptop GPU
Create a dozen background tracks and a sung title theme
Run the turbo model with the 0.6B planner in the Gradio app; each track takes seconds, the MIT licence permits shipping them in the game, and repaint fixes a weak bridge without regenerating the rest.
A producer working on a Mac
Teach the model their own style and finish tracks in a DAW
Use the macOS launch script (MLX), train a LoRA on 8 of their own songs, then split the output into stems for mixing; the README's figure of about an hour for training is for a 12 GB RTX 3090, and Mac timing is not published.
The backend team of a SaaS video editor
Offer users generated soundtracks without per-song API fees
Run acestep-api, the bundled REST server, on GPU instances and batch up to 8 songs per call; with no licence fee, the cost is GPU time alone.
Before you commit
- Setup
- Install with uv, then start the Gradio app on port 7860 or the REST API on port 8001; launch scripts cover Windows, Linux and macOS, and a portable Mac package is offered. Keep those ports off the public internet unless you add access control.
- Integrations
- Python inference API, REST API, Gradio UI and ComfyUI nodes, plus the community Side-Step toolkit for advanced LoRA training.
- Data and privacy
- Self-hosted, so prompts and audio stay local unless you use acemusic.ai. The team says training used licensed music, royalty-free and public-domain music, and synthetic audio rendered from MIDI.
- Cost at scale
- No licence fees at any scale; the cost is GPU time, with up to 8 songs generated per batch.
- Limits
- 10 seconds to 10 minutes per song; at 6 GB of VRAM or less the planner model is off by default; XL needs 12 GB with offloading and quantisation or 20 GB without; covers change the melody.
What’s good
- MIT licence; commercial use allowed
- Turbo model runs in under 4 GB of VRAM; a song in under 10 seconds on an RTX 3090
- 10 seconds to 10 minutes, lyrics in 50+ languages
- Repaint, stems, layers, vocal-to-accompaniment and LoRA training
- Runs on CUDA, Apple Silicon, AMD ROCm, Intel and CPU
What’s not
- Lower WildSongBench scores than MiniMax Music 3 and YuE2
- Covers do not keep the source melody
- XL quality needs 12 to 20 GB of VRAM
- Training data described only by category
- Core model dates from January 2026
Best for
Developers, game studios and video makers who need commercially usable songs from a model that runs on an ordinary GPU or a Mac.
ACE-Step 1.5 alternatives
ACE-Step 1.5: common questions
- Is ACE-Step 1.5 worth it?
- ACE-Step 1.5 is the open music model to start with: MIT-licensed, fast on a modest GPU and full of editing tools, even if the newest open rivals sound better. We rate it 4.3 out of 5.
- How much does ACE-Step 1.5 cost?
- Free under the MIT licence, commercial use included. The team's hosted app at acemusic.ai is free. The turbo model self-hosts on a GPU with under 4 GB of VRAM.
- What are the downsides of ACE-Step 1.5?
- Lower WildSongBench scores than MiniMax Music 3 and YuE2; covers do not keep the source melody; xL quality needs 12 to 20 GB of VRAM; training data described only by category; core model dates from January 2026.
- Who is ACE-Step 1.5 best for?
- Developers, game studios and video makers who need commercially usable songs from a model that runs on an ordinary GPU or a Mac.
- How much does ACE-Step 1.5 cost for a team or in production?
- No licence fees at any scale; the cost is GPU time, with up to 8 songs generated per batch.
- Is ACE-Step 1.5 safe for company data?
- Self-hosted, so prompts and audio stay local unless you use acemusic.ai. The team says training used licensed music, royalty-free and public-domain music, and synthetic audio rendered from MIDI.
- What are the best ACE-Step 1.5 alternatives?
- The closest alternatives we have reviewed are MiniMax Music 3, YuE2-3B and Stable Audio 3.0 Medium.
Answers checked Nov 7, 2026.
Make ACE-Step 1.5? Show this rating on your site
