MiniMax has officially released the weights for MiniMax-Music3, along with an online demo and a ComfyUI workflow. The model can take lyrics and a music description and generate a complete song up to about five minutes long, outputting 32kHz, 16-bit stereo audio.
Music3's headline feature is controllability. You can structure songs directly with tags like [Intro], [Verse], [Chorus], [Bridge], [Solo], and [Outro] in the lyrics. The prompt can also specify genre, BPM, key, mood shifts, vocal timbre, harmonies, instrumentation, and the arrangement for each section.
And this time, you can actually download the model and run it locally. MiniMax says full precision needs under 24GB of VRAM, around 22GB with CPU offloading, and if you're willing to load layers sequentially, it can even squeeze into an 8GB GPU. ComfyUI has native support, and there's an INT8 version for low-VRAM setups.
Among open-weight music models, Music3 is already being called one of the strongest available. After H3, MiniMax has now dropped another heavyweight generative model directly into the wild.
