MiniMax Unveils Music3, a Music Model Capable of Generating Songs Up to Five Minutes Long From Lyrics and Descriptions

MiniMax Unveils Music3, a Music Model Capable of Generating Songs Up to Five Minutes Long From Lyrics and Descriptions

MiniMax today unveiled MiniMax-Music3, a music-generation model that turns lyrics and musical descriptions into complete songs of up to five minutes, output as 32kHz, 16-bit stereo WAV files.

The model is underpinned by a hierarchical autoregressive architecture in which two components operate in concert. An 8B global language model, initialized from Qwen3-8B, predicts the first RVQ codebook frame by frame, governing long-range semantics and structural evolution; its embedding and output layers were first adapted to music-semantic tokens before joint modeling of all codebooks with the local model. A 0.6B local language model then predicts the remaining acoustic codebooks within each frame, recovering fine-grained acoustic detail. Structure is set by the large model; sonic texture is filled in by the small one—a clearly delineated division of labor. [IMAGE:0]

MiniMax states that the model sustains musical theme, rhythm, vocal identity, and arrangement progression across extended audio, preserving each structural section—intro, verse, pre-chorus, chorus, bridge, instrumental interlude, and outro. A model page has been published on GitHub, with generated song samples made available for listening.