MiniMax took the wraps off a new music model on Monday that turns two inputs into a finished five-minute track in one pass. The release is open-weights, with inference code and three documented serving paths available from day one.
The first input is lyrics carrying section tags such as verse, chorus, and bridge; the second is a structured caption that spells out vocals and arrangement. Output comes back as 32 kHz 16-bit stereo audio.
On the output side, the design skips token decoding altogether. The two models’ final hidden states feed a flow-matching module, and a compact VAE turns the result into audio. An 8B global language model holds long-range musical structure while a 0.6B local model fills in acoustic detail frame by frame.
At training time, audio passes through eight stacked quantization layers that separate musical structure from acoustic detail.
The company documents three deployment routes. One, SGLang-Omni, is a reference server spread across two CUDA GPUs; another is a diffusers pipeline that needs under 24 GB of VRAM, dropping to roughly 8 GB with offloading; a third is a ComfyUI template with repacked weights. One loose end: the model card lists Qwen3-8B as the base, while MiniMax’s own research post says Qwen3.5-8B.
