ACE-Step is an open AI model for making music: from song lyrics and style tags it produces a track with vocals. The gpu.nz template runs a web interface and an API, so you can work in the browser or from your own code. The song is generated on the GPU, and you pay by the minute while the server runs.
In short
ACE-Step runs on cards with 12 GB or more. The most affordable suitable card right now is the RTX 3060 12 GB from 6.95 ₽ per hour, billed by the minute.
Updated 13 October 2026
The minimum is 12 GB of video memory, as in the template, which covers short and medium tracks. For long songs, several takes in a row or API work, take 16 GB or more: an RTX 4090 24 GB gives plenty of room. For trial tracks an RTX 3060 12 GB is enough.
Prices and availability come from the live catalog and update automatically.
Mark the lyrics with structure tags: [verse] for a verse, [chorus] for a chorus and [bridge] for a bridge. The tags help the model work out what is sung where. You can draft the lyrics in a chat with a language model, for example on the Open WebUI + Ollama template. A short lyric with tags looks like this:
[verse]
A night avenue, lamps in the fog,
I walk home while the city sleeps.
[chorus]
Don't go out, light, while I am here,
let my road become a song.The style is set with a line of tags, usually in English:
synthwave, female vocal, 110 BPM, warm analog synthA full song is usually ready in a minute or two, and the first start is slower because the models load onto the server's disk. Check the lyrics and the style on one or two generations first, then run the full versions, so less time goes on unsuccessful takes. Another way to avoid overpaying is to stop the server between sessions, see how to spend less on GPU rental.
For Russian lyrics, check the pronunciation on the first generations, since quality depends on the model. If a word sounds wrong, replace it with a simpler synonym.
Finished tracks download from the interface, and for batch work it is easier to take them through the API. If you plan to work with the model for several days, do not delete the server: stopping it keeps the models on the disk without using GPU minutes, while downloading them again takes time. A large batch of tracks through the API suits an interruptible server if you download the files as they are ready: see interruptible servers.