MiniMax unveils Music 3.0 full-song model
MiniMax's Music 3.0 generates complete songs of up to five minutes from a single prompt, and the lab bills it as open-weights, though the post gives no release details or benchmarks.
MiniMax announced Music 3.0, a music generation model that takes a creative concept and optional lyrics and produces a finished song, up to five minutes long, in a single generation. The lab says it rebuilt the whole pipeline to fix common problems in AI music: requested instruments dropping out partway through, emotional tone fading, and vocal styles that hold for only one section.
The architecture has three parts. An eight-layer residual vector quantization (RVQ) tokenizer puts song structure in its first layer and acoustic detail in the other seven. A "Hybrid-LM" pairs an 8B global model, initialized from Qwen3.5-8B, that predicts semantic tokens frame by frame, with a 0.6B local model, trained from scratch, that predicts acoustic tokens within each frame. Instead of passing discrete tokens straight to a decoder, the system combines the continuous hidden states of both models to condition a 2.4B flow-matching module. A 123M Flow-VAE then decodes the audio. Altogether the components come to roughly 11B parameters.
For conditioning, MiniMax uses "Structured Captions," time-stamped descriptions of genre, tempo, key, emotional arc, when instruments enter and exit, and how vocal delivery changes by section. A template-based prompt enhancement system expands short user prompts into these detailed captions. Lyrics can also carry section tags such as [verse] and [chorus].
The title calls the model open-weights, but the post says nothing about licensing, where to download it, or when it will be released. It also includes no quantitative evaluations or comparisons with other music models.