MiniMax Releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption

Details

MiniMax-Music3 takes two structured inputs — tagged lyrics and a detailed music description — and outputs up to five minutes of 32 kHz, 16-bit stereo audio in a single pass. Unlike models that generate short clips, Music3 targets full-song output from the start, making it directly useful for game audio, creator tools, and music-tech SaaS. The architecture pairs a hierarchical autoregressive stack with a continuous synthesis path. A global 8B LLM (initialized from Qwen3/3.5-8B — the exact checkpoint is unsettled per the source) predicts semantic RVQ codebook frames for long-range musical structure, while a 0.6B local LLM handles acoustic detail per frame. Rather than decoding from discrete tokens, final hidden states from both LLMs feed a 2.4B flow-matching module backed by a 123M Flow-VAE, avoiding the quality ceiling that discrete-only systems typically hit. The MiniMax-Music3 Community License permits commercial use with prominent attribution required. Organizations whose aggregate annual revenue from Music3-powered products exceeds 0M must obtain separate written authorization from MiniMax — a threshold that screens out large enterprises without restricting indie studios or mid-market teams. Third-party hosting platforms must also implement safeguards against infringing outputs. Weights, inference code, and three documented serving paths shipped on day one. This is a deployable release, not a research preview. Practical targets include game development, advertising, UGC video scoring, e-learning, and fitness apps.