Suno charges $30/month for its Premier plan. Udio has similar pricing. Both keep your prompts, generated audio, and usage patterns on their servers. If you want to generate AI music without the subscription fees or privacy tradeoffs, ACE-Step 1.5 runs entirely on your own machine.
Released in late January 2026 by ACE Studio and StepFun, ACE-Step 1.5 is an open-source music generation model that matches or exceeds commercial alternatives on standard benchmarks. It generates full songs in under 10 seconds on consumer GPUs, needs less than 4GB of VRAM for basic operation, and supports 50+ languages.
What You’ll Need
Minimum requirements:
- 4GB VRAM (GPU with CUDA, Apple Silicon, AMD ROCm, or Intel XPU)
- 8GB system RAM
- Python 3.11 or 3.12
- ~10GB disk space for models
Recommended for best results:
- 8-16GB VRAM for better quality
- 16GB+ VRAM for the full 4B language model
- SSD for faster model loading
ACE-Step runs on Windows, macOS, and Linux. Apple Silicon Macs use the MLX backend, AMD cards use ROCm, and Intel GPUs use XPU drivers.
Quick Installation (3 Minutes)
The project uses uv for dependency management, which handles everything automatically.
Step 1: Install uv
On macOS or Linux:
curl -LsSf https://astral.sh/uv/install.sh | shOn Windows (PowerShell):
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"Step 2: Clone and install
git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv syncStep 3: Launch
uv run acestepOpen http://localhost:7860 in your browser. Models download automatically on first run (~3-8GB depending on your VRAM tier).
Alternative: Portable Packages
If you don’t want to deal with command lines, ACE-Step offers portable packages for Windows and macOS with pre-installed dependencies. Download from the releases page, extract, and run the startup script.
Using ACE-Step
The Gradio interface has two main modes: Generation and Editing.
Generating New Songs
- Enter a text prompt describing the song style, mood, and genre
- Add lyrics if you want vocals (the model supports 50+ languages)
- Set duration (up to 10 minutes)
- Click Generate
Example prompt:
Ambient electronic, dreamy synthesizers, 120 BPM,
minor key, layered pads, subtle percussionThe interface auto-selects the optimal model configuration for your GPU. On an RTX 3090, expect generation in about 8-10 seconds. On an RTX 4090, it’s closer to 5 seconds.
Editing and Variations
ACE-Step includes tools for:
- Cover generation: Create variations of existing songs
- Repainting: Modify specific sections while keeping the rest
- Track separation: Extract vocals, drums, bass, and other stems
- Vocal-to-BGM conversion: Strip vocals and generate background music
Training Custom Styles (LoRA)
If you have a specific sound you want to replicate, ACE-Step supports LoRA fine-tuning from just a few reference songs. This is useful for matching a particular artist’s style or creating consistent output for a project.
Model Tiers by VRAM
The language model component affects output quality. ACE-Step automatically selects based on your VRAM:
| VRAM | LM Model | Quality |
|---|---|---|
| ≤6GB | DiT only (no LM) | Basic |
| 6-8GB | 0.6B parameter | Good |
| 8-16GB | 0.6B-1.7B parameter | Better |
| 16-24GB | 1.7B parameter | Very good |
| ≥24GB | 4B parameter | Optimal |
Even the smallest tier produces usable music. The larger models improve lyric adherence, style consistency, and overall coherence.
How It Compares to Suno
On SongEval benchmarks, ACE-Step 1.5 outperforms Suno v5 on overall music quality scores. Independent human evaluations place it between Suno v4.5 and v5 in subjective quality.
Where Suno still leads:
- Slightly better style and lyric alignment
- More polished web interface
- No setup required
Where ACE-Step wins:
- Free forever (MIT license)
- Complete privacy—nothing leaves your machine
- Unlimited generations
- No watermarks or usage restrictions
- Full commercial rights
- LoRA customization
At Suno’s Premier pricing ($30/month), you’d spend $360/year for 2,000 songs per month with usage caps. ACE-Step costs $0 after your electricity bill.
API Access for Automation
For programmatic use, start the REST API:
uv run acestep-apiThe API runs at http://localhost:8001 with endpoints for generation, editing, and batch processing. You can generate up to 8 songs simultaneously if your VRAM allows.
Privacy Considerations
Running locally means:
- No prompts sent to external servers
- No generated audio stored in the cloud
- No usage analytics or tracking
- No training on your outputs
If you’re creating music for commercial projects or simply prefer not sharing your creative process with AI companies, this matters.
Troubleshooting
“CUDA out of memory”: Lower the batch size or switch to a smaller LM model tier in the settings.
Slow generation on AMD: ROCm support varies by card. Check the compatibility list for known issues.
macOS MLX errors: Make sure you’re on macOS 14+ with Apple Silicon. Intel Macs aren’t supported.
Models not downloading: Check your internet connection and disk space. Models range from 3-8GB.
What This Means
Commercial AI music services solved a genuine problem—making music generation accessible without technical expertise. But they also introduced subscription fatigue and privacy concerns that weren’t inherent to the technology.
ACE-Step 1.5 demonstrates that commercial-grade music generation can run locally on hardware many creators already own. The 4GB VRAM floor puts it within reach of most discrete GPUs from the last five years.
The tradeoff is setup complexity. Suno’s advantage is that you don’t need to install anything. If you’re comfortable with command lines and have a capable GPU, that tradeoff likely favors ACE-Step.
What You Can Do
- Check if your GPU meets the minimum 4GB VRAM requirement
- Install using the quick setup above
- Start with simple prompts to understand the model’s capabilities
- Explore LoRA training if you need consistent output styles
- Consider the musician’s guide for production-quality tips
For more self-hosting guides that replace paid cloud services, check our local-ai category.