MiniMax-Music3: Complete 5-Minute Songs from Two Lines of Lyrics, Intro to Outro

MiniMax Music 3 generates complete songs up to 5 minutes long, conditioned on lyrics and detailed music descriptions. It maintains the structural flow of the song—including intro, bass, pre-chorus, chorus, bridge, and outro—while consistently expressing vocal identity, rhythm, and arrangement evolution even in long sequences.

The model uses a hybrid architecture combining an 8B-parameter Global LLM and a 0.6B-parameter Local LLM. The Global LLM handles the long-term structure and meaning of the song, while the Local LLM processes fine-grained acoustic information at the frame level. By fusing the hidden states of these two LLMs, it outputs 32kHz, 16-bit stereo WAV audio via Flow Matching and Flow-VAE.

The inputs are lyrics and a music description. The lyrics can include section tags such as [Verse] and [Chorus], while the music description allows specifying details like genre, BPM, key, vocal style, and instrumentation. Using structured captions enables precise control over not only the overall style of the song but also its musical development over time.

Inference is possible in a CUDA environment, and currently only non-streaming generation is supported. Text prompts are limited to 5,000 tokens, and audio generation is limited to 9,000 acoustic frames. Section tags and music descriptions serve as generative control means, and not all requested details may be accurately reflected.