Magenta RealTime vs ACE-Step
These are the two open models you can actually ship without a licensing argument. They answer completely different questions: one generates while you listen, the other writes finished songs.
Side by side
| Magenta RealTime | ACE-Step | |
|---|---|---|
| Generation | Faster than playback — streams | ~34× real time, but batch |
| Interaction | Steer it while it plays | Prompt, wait, receive |
| Vocals | No | Yes |
| Length | Short-form | Minutes |
| Licence (weights) | CC-BY-4.0 — attribution only | Apache-2.0 |
| Primary platform | Apple Silicon (MLX) | CUDA, plus Apple Silicon |
| NVIDIA | Offline inference only | Full support |
| Sizes | 230M and 2.4B | 3.5B |
Where they actually differ
"Fast" and "real-time" are not the same claim
ACE-Step is very fast — a minute of audio in under two seconds on a 4090 — but you still submit a job and collect a result. Magenta RealTime generates faster than the audio plays, which means you can change the prompt mid-stream and hear it respond. That is a different category of tool: one is a renderer, the other is an instrument.
The platform story is inverted
Almost every model here treats Apple Silicon as a compromise. Magenta RealTime treats it as the target — it uses MLX, and real-time streaming requires M-series hardware, with NVIDIA GPUs limited to offline inference. ACE-Step is the conventional way round: CUDA-first, with usable Apple Silicon support. Your hardware may decide this outright.
Both licences are clean, in different ways
CC-BY-4.0 asks for attribution and nothing else. Apache-2.0 asks for essentially nothing. These are the two most permissive positions in open music generation, and either is safe to build on — which cannot be said of MusicGen, YuE2, RAVE, MAGNeT or AudioGen.
Scope is the real divider
Magenta RealTime generates short-form audio. It does not write verses and choruses, and it does not sing. ACE-Step produces a structured track with lyrics you supplied. If the deliverable is a song, there is no contest.
Which should you choose?
Choose Magenta RealTime when…
- You are performing live or building something interactive
- You want to steer generation as it plays
- You are on a Mac — including an Air
- You want the most permissive weights available
Choose ACE-Step when…
- You need finished songs with vocals and structure
- You are on an NVIDIA GPU
- You want editing — retake, repaint, extend
- You need output measured in minutes