ACE-Step 1.5, from ACE Studio and StepFun, came out in January 2026 with a paper on arXiv (2602.00744). It works in two stages, a language model that plans the song and a diffusion transformer that renders it, and it's light enough to run on a modest consumer GPU.
What you get running your own
- Songs from 10 seconds to 10 minutes, with vocals and your own lyrics.
- Editing tools: covers, repainting a section, generating an accompaniment from a vocal.
- Your own style: LoRA adapters let you train it toward a sound.
- No credits, no queue, no download cap. It runs on your machine.
What you give up is the product around the model: a polished interface, managed updates, a shared library, someone else's servers, and the support and terms of a company. You also take on the licence question yourself. Check the model licence and think about the training data before you release anything made with it. Coverage describes it as Apache-2.0, but confirm that on the project's own repository.
How Suno compares
Suno is the hosted version of the same idea: v6 up to eight minutes, Replace Section, Extend, Cover, stems, Voices, custom models trained on six or more of your tracks, Studio on Premier, and published commercial terms on paid plans for songs downloaded within the monthly allowance. Suno says v6 was trained on licensed material from Warner, BMG and Believe.
Prompting is the shared skill
Whichever model you run, the prompt habits carry over: name the genre first, then tempo and feel; describe parts, not vibes; put a direction on every section; write out every repeat; end on a cue. Our v6 prompt guide is written for Suno, but most of it applies to any model that takes lyrics with section tags, and the free lyric checker doesn't care which model you paste for.
The bottom line
Tinkerers, researchers and anyone who wants no caps and full control: ACE-Step. Songwriters who want to spend their time on the song rather than the setup: Suno.