MiniMax open-sources Music3 music generation model producing 5-minute complete songs under Apache-2.0-compatible terms
MiniMax released Music3 on August 14, a music generation model using a hierarchical autoregressive architecture that takes lyrics and music descriptions to produce songs up to 5 minutes at 32kHz/16-bit stereo WAV, with weights open-sourced on Hugging Face and ModelScope and Day-0 ComfyUI support.
On August 14, MiniMax released Music3, a music generation model that takes lyrics and music descriptions to produce complete songs up to 5 minutes long, outputting 32kHz, 16-bit stereo WAV. The model uses a hierarchical autoregressive architecture: an 8B-parameter Global LLM handles long-range semantic and structural modeling, a 0.6B-parameter Local LLM restores fine-grained acoustic detail, and a 2.4B flow-matching module plus a 123M Flow-VAE decoder handle synthesis.
MiniMax-Music3 maintains musical theme, rhythm, vocal identity and arrangement progression across long audio. Weights are available on Hugging Face and ModelScope. The open-weight license is CC BY-NC 4.0, with commercial use requiring written authorization above USD 20 million in annual revenue. Three deployment paths are shipped: SGLang-Omni, Diffusers and ComfyUI. ComfyUI 0.33.0 added native Music 3 support and Comfy Cloud also provides access.
The release lands in a dense Chinese model launch week. On the same day, DeepSeek opened its Harness agent framework preview, Zhipu GLM-5.3 set new open-source records purely through post-training, and Google released Gemini 3.7 Flash. MiniMax's contribution was extending AI-composed song length to a full 5-minute structured track with intro, verse, chorus and bridge.