MiniMax Music 3 AI Music Generator

No, but the song does need words. MiniMax Music 3 is conditioned on lyrics, so the composer asks for them — and if you would rather not write them, it will generate a set from your brief that you can edit line by line before you generate the song. For music with no singing, tag the sections [Instrumental] or [Solo] instead.

Up to five minutes in a single generation, set on a slider anywhere between a few seconds and 300. The model was built for full-length songs and holds its theme, tempo and vocal identity across that whole span, which is what separates it from clip generators that top out around thirty seconds.

Credits, and the number scales with the duration you ask for — a short sketch costs a fraction of a five-minute track. The Create button shows the exact cost for your current settings before you spend anything, and signing up to Musci3 gives you 15 credits to start with.

Yes. Creating an account gives you 15 free credits — enough for a sixty-second song or a few shorter sketches — and no card is needed to use them.

It is the brief, written in the shape MiniMax recommends: global metadata (genre, BPM, key, mood), vocal details (who is singing and how), and arrangement (which instruments, entering when). A plain sentence works too, but the structured version gives you far more control over what comes back. Every example on this page is written that way.

MiniMax publishes Music 3 under its own community licence, which permits commercial use with attribution and requires a separate agreement from MiniMax only above a high revenue threshold. That licence, not this page, is the authority — read it, and check your plan here, before you build a business on the output.

MiniMax Music 3 is built around one thing: a complete, structurally coherent song from a brief and lyrics, in one pass, up to five minutes. Suno gives you a wider toolkit around a song — two takes per generation, reference audio, saved voices, stem separation, covers and extensions. Musci3 runs both, and the model select in the composer switches between them.

Not from Music 3 — it returns a finished mix. Stem separation is part of the Suno toolkit here, so switch the model in the composer if you need a song split into its parts.

The weights are published on Hugging Face under the MiniMax-Music3 community licence, so you can run the model yourself if you have the hardware for it — reviewers have reported needing tens of gigabytes of GPU memory. Musci3 runs a hosted copy so you do not need any.

Usually about a minute for a short song, and longer as you ask for more duration. You do not have to wait on the page: the job keeps running and the finished track appears in your library.

On MiniMax Music 3, text is the only input: you write a description of the music and — unless you tag the sections instrumental — the lyrics, and the model returns a finished recording. So text to music and text to song both describe it exactly, and Musci3 will write the words from your description if you would rather not. Suno, the other model here, additionally takes reference audio and saved voices.