MiniMax Open Source Music3: Generates a single full-length song of up to approximately 5 minutes, outputting 32kHz, 16-bit stereo audio

According to Perceiving Beating monitoring, MiniMax has officially released the MiniMax-Music3 weights, and simultaneously launched an online demo and ComfyUI workflow. It can generate a complete song of up to about 5 minutes in length at once based on lyrics and music description, outputting 32kHz, 16-bit stereo audio.

Music3's biggest feature is controllability. Lyrics can be directly arranged into song structures using tags such as [Intro], [Verse], [Chorus], [Bridge], [Solo], [Outro]. Cue words can also specify the musical style, BPM, key, emotional changes, vocal timbre, harmony, instruments, and arrangement of each paragraph.

Moreover, this time it's really downloadable to run the model locally. The official statement indicates that full precision requires less than 24GB of VRAM, about 22GB for CPU offloading, and can even be squeezed into an 8GB graphics card through further layer-by-layer loading. ComfyUI has native support and also provides an INT8 model for low VRAM usage.

In the open-weight music model, Music3 has been considered by many users as currently one of the most powerful ones. Following H3, MiniMax has once again released a set of heavyweight generation models directly.