MiniMax-Music3

MiniMax Music 3 (MiniMax-Music3) is a music generation model developed by the Chinese company MiniMax. Conditioned on lyrics (optionally with section tags such as [Verse] or [Chorus]) and a style description, it produces complete, structurally coherent songs of up to five minutes with vocals and evolving arrangements. The output is 32 kHz, 16-bit stereo WAV audio.

The model combines a hierarchical autoregressive architecture with a synthesis system based on Flow Matching and Flow-VAE. An 8B Global LLM (initialized from Qwen3-8B) models the song's long-range structure, while a 0.6B Local LLM restores frame-level acoustic detail. The synthesis module (2.4B Flow Matching and a 123M Flow-VAE decoder) generates a continuous audio representation. The training tokenizer uses eight layers of Residual Vector Quantization (RVQ).

The model weights are publicly available on Hugging Face under the MiniMax-Music3 Community License. Inference requires an NVIDIA GPU (CUDA) and is supported by SGLang-Omni, diffusers and ComfyUI, among others. Only non-streaming generation is supported and the tokenized text prompt is limited to 5,000 tokens.