What Happened
MiniMax introduced Music 3.0 on August 13, 2026. The model generates complete songs from a creative description and optional lyrics, with support for pieces up to five minutes long. This entry dates the model announcement; it does not independently establish the first public download date of every associated checkpoint.
Why It Matters
The release reflects progress from short audio samples toward generation that maintains a musical idea across a complete composition. Its technical design explicitly separates song-scale organization from fine acoustic detail, making it relevant to the history of long-form generative media.
Technical Details
Music 3.0 combines a global-local Hybrid-LM with an eight-layer residual-vector-quantization representation. The global component models musical structure, while the local component predicts frame-level acoustics. Their hidden states condition flow matching and a Flow-VAE synthesis stage. MiniMax presents the release as open weights; specific licensing and deployment conditions require the associated checkpoint documentation.