Create original music tracks from text descriptions using ACE-Step, running on your own GPU with no limits.

What is ACE-Step?

ACE-Step is an open-source text-to-music model that generates full audio tracks from natural language descriptions. Tell it "upbeat jazz with piano and saxophone" and it produces a listenable track in under a minute on consumer hardware.

Requirements

GPU: 10 GB+ VRAM recommended
Storage: ~4 GB for model weights
Platform: Linux (native ROCm/CUDA) or Windows (CUDA)

Installation

Find ACE-Step in the Thumper-Run catalog and click Install. The install pipeline handles the Python environment, model downloads, and GPU configuration. On AMD GPUs, Thumper automatically sets MIOPEN_FIND_MODE=2 and uses BF16 precision to avoid the NaN issues described in our AMD GPU guide.

Your First Track

Launch ACE-Step through Thumper
Enter a prompt: "ambient electronic, slow tempo, atmospheric pads, rain sounds"
Set duration (10–60 seconds for your first test)
Click Generate

A 30-second track generates in about 9 seconds on an RTX 4060 Ti. Longer tracks scale linearly.

Prompt Tips

Specify genre and instruments: "orchestral, strings, french horn, epic" works better than "epic music"
Include tempo: slow, medium, upbeat, fast
Add mood: melancholic, triumphant, mysterious, playful
Describe texture: lush, sparse, layered, minimalist

Performance Notes

For songs over 60 seconds, ACE-Step switches to tiled VAE decoding to avoid VRAM overflow. This is slower but necessary.
First-run on AMD includes a MIOpen warmup phase (5–15 minutes). Subsequent runs use cached kernel tunings.
Close other GPU apps during generation to maximize available VRAM.

See the VRAM requirements guide for detailed hardware recommendations.

Use Cases

YouTube/TikTok background music without licensing fees
Game development prototyping
Podcast intros and outros
Combine with [AI video](/blog/ai-video-generation-local-ltx-hunyuan) for fully AI-generated media

Ready to try it? Download Thumper-Run free →