MiniMax has launched the music generation model MiniMax-Music3. By inputting lyrics and a musical description, it can generate a complete song up to 5 minutes long, with the output being a 32kHz, 16-bit stereo WAV file.
The model is supported by a hierarchical autoregressive architecture, where two models each have their own responsibilities and work together. The global language model (Global LLM) has a parameter scale of 8B, and it predicts the first RVQ codebook frame by frame, specifically responsible for modeling the long-term semantics and structural changes of the song. It is initialized from Qwen3-8B. During training, the embedding layer and output layer are adapted to musical semantic tokens first, then jointly model all codebooks with the local model. The local language model (Local LLM) has only 0.6B parameters and is responsible for predicting the remaining acoustic codebooks in each frame, gradually filling in the fine-grained acoustic information. One large model manages the structural skeleton, while a small model fills in the sound details. Their division of labor is very clear.
The official stated that the model can firmly maintain the musical theme, rhythm, vocal identity, and progression of the arrangement in long audio, without missing any structure such as prelude, verse, pre-chorus, chorus, bridge, instrumental interlude, and coda. The official has already released the model page on GitHub and provided sample songs for listening.
