Capable of generating full 5-minute studio-grade tracks with paragraph-level precision control. Versions from Music 1.5 to 2.5+.

Music 3.0 (API July 16, 2026; weights on HuggingFace August 13, 2026) is the first open-weight release in the line, under the MiniMax-Music3 Community License. It generates complete songs of up to five minutes as 32 kHz stereo from lyrics plus a music description. The technical blog describes three parts: an eight-layer RVQ tokenizer whose first layer carries structure and whose other seven carry acoustic residuals; a Hybrid-LM in which an 8B Global LLM initialized from Qwen3.5-8B predicts frame-level semantic tokens and a randomly initialized 0.6B Local LLM predicts the within-frame acoustic tokens along the depth axis; and a synthesis stack in which fused hidden states from both LLMs condition a 2.4B flow-matching module decoded by a 123M Flow-VAE, rather than feeding discrete tokens to a decoder. Training runs global alignment first, then full-parameter joint training of both LLMs.

Model Details

License MiniMax-Music3 Community License
audiomusicgeneration

Notes

Music-01 launched August 2024; Music 2.0 October 2025; Music 2.5 early 2026; Music 3.0 July 2026 (open weights August 13, 2026).