MiniMax Music 3

MiniMax Music 3 generates end-to-end songs up to five minutes long in a single pass by separating structural lyrics from stylistic audio prompts. Driven by an 8B-parameter Global LLM for song-level arrangement and a 0.6B local acoustic model, it produces complete multi-verse tracks with natural breathing pauses, harmonic vocal backing, and dynamic instrumental shifts.

Enter bracketed section tags such as [Intro], [Verse], [Chorus], [Bridge], [Solo], [Instrumental], and [Outro] on their own dedicated lines above each lyric block. Do not place lyrics on the same line as the bracket tag, and keep rhyming cadence consistent across lines so the vocal synthesis phrases naturally without slurring.

Keep the guidance scale between 1.5 and 2.5 (default 1.7) and set inference steps near 30. Raising guidance above 5.0 can cause acoustic clipping and robotic vocal timbre, while increasing steps past 50 provides diminishing returns and risks introducing phase artifacts into the mix.

MiniMax Music 3 excels at contemporary pop, indie rock, electronic dance music, cinematic soundtrack cues, and structured commercial jingles with clear vocal leads. It handles multi-verse progressions and acoustic layering with balanced stereo production.

Avoid MiniMax Music 3 when you require isolated multitrack DAW stems, sample-accurate BPM grid lock for beatmatching, or complex non-Western polyrhythmic traditions like Latin micro-rhythms and intricate African drumming. It produces rendered stereo masters rather than separable instrument tracks.