MiniMax Music 3.0 highlights structure: write how the arrangement changes

MiniMax introduced Music 3.0 in an August 13 research post. It describes complete-song generation from a concept and optional lyrics, with structured descriptions of how musical elements change across a piece. This is an older release retained as context for Chinese AI music coverage, outside our current 30-day window.

Describe a change, not a bag of instruments

A global description tells the model what belongs in the song. A section description tells it when a choice becomes important. For a first exercise, keep the global sound simple: one genre, two instruments and one production texture.

Then write three contrasting states. The opening could use piano alone; the verse could add a low supporting part; the chorus could widen the arrangement. Do not add a different genre to each section just to create contrast. Changes in density, register and delivery give you other variables to test.

An original sketch is: “A restrained Mandarin folk-pop song. Piano stays at the center. The first verse is close and dry; the final chorus adds a sustained string layer, then returns to piano for the ending.” Treat this as a writing example, not proof that a particular model will obey every transition.

Record the surface you used

The research post is not a substitute for an API contract or a model license. If you plan to use hosted generation or local weights, follow the provider’s current links and verify that the specific model, access method and requirements match your environment. Do not derive an endpoint or a parameter name from the announcement title.

In MusicPrompt Studio, put global choices in the sound selector and local directions on the relevant sections. The selector tutorial gives a complete exercise; a saved blueprint is the evidence you need when comparing later generations.