MiniMax-Music3 combines an 8B Global LLM and continuous hidden-state synthesis to turn lyrics and captions into 32 kHz stereo tracks in a single run.
MiniMax officially released MiniMax-Music3, an open-weights text-to-music generation model capable of creating complete five-minute songs in a single generation. The model generates output as 32 kHz, 16-bit stereo WAV audio.
The system accepts two separate inputs: lyrics containing structural section tags and a structured caption describing genre, tempo, instrumentation, and vocal details. The architecture utilises a Hybrid-LM framework combining an 8B Global LLM for long-range structural modelling with a 0.6B Local LLM for frame-level acoustic details.
The synthesis pipeline fuses the continuous hidden states of the LLMs to condition a 2.4B flow-matching module and a 123M Flow-VAE decoder, bypassing the discrete tokenizer decoder during inference. The training tokenizer uses an eight-layer residual vector quantisation (RVQ) scheme, featuring a 16,384-entry semantic codebook and seven 1,024-entry acoustic codebooks.
MiniMax published complete model weights (totaling 57.4 GB), inference code, and three documented deployment paths, including SGLang-Omni, Diffusers, and ComfyUI. The reference server setup uses SGLang-Omni across two CUDA GPUs, while the Diffusers pipeline can run within 24 GB of VRAM or down to 8 GB using group offloading.
The model is released under the MiniMax-Music3 Community Licence, which allows commercial use with prominent UI attribution but requires custom licensing for organisations earning over $20 million in annual revenue.
