MiniMax Launches MiniMax-Music3: Open-Weights Music Generator Creating Full Five-Minute Songs From Lyrics and Descriptions

MiniMax released an open-weights music generation model called MiniMax-Music3 that converts text input—lyrics with section tags and detailed music descriptions—into complete five-minute songs. The model pairs a 8-billion-parameter global language model with a smaller 600-million-parameter local model, combined with advanced flow-matching synthesis technology. What sets this release apart is that MiniMax shipped production-ready weights, inference code, and three documented serving implementations on day one, making the model immediately deployable. The system generates 32 kHz, 16-bit stereo audio using an eight-layer vector quantization system where the primary layer captures core musical structure and seven additional layers encode acoustic details. The synthesis pipeline uses flow matching instead of discrete token decoding, directing a 2.4-billion-parameter flow module. Three deployment options are documented: SGLang-Omni as the reference server needing two GPUs, the diffusers library supporting configurations from 8 GB to 24 GB of memory, and ComfyUI with optimized weights.