MiniMax Music 3

Complete songs rather than loops are the output of MiniMax Music 3: given lyrics annotated with section tags and a structured caption describing genre, vocal character, and arrangement, it renders up to five minutes of 32 kHz 16-bit stereo audio with coherent verse and chorus structure. The open-weights release exposes the full stack: an 8B global language model plans long-range musical structure, a 0.6B local model fills in acoustic detail over an 8-layer residual vector quantization tokenizer, and a 2.4B Flow Matching network with a Flow-VAE decoder synthesizes the waveform. The reference pipeline splits those stages across two CUDA GPUs, but CPU offloading brings it down to a single 8 GB card with 24 GB or more recommended, and ComfyUI shipped official workflow support at launch. Weights live on Hugging Face under the MiniMax-Music3 Community License, which permits commercial use below 20 million dollars of annual revenue with attribution in the product UI. Released in August 2026, it is the strongest open lyric-to-song system yet from a major lab.