MiniMax opens Music 3.0 weights: 5-min vocal songs, run local
Commercial use is free under $20M revenue, no country carve-out this time, and GGUF quants already fit it in 8GB. Plus: Google open-sources an FHE compiler.
Copy markdown
Five-minute songs, not clips
Music 3.0 writes complete tracks up to five minutes — intro, verses, chorus, bridge, outro — with expressive vocals at 32kHz stereo. You feed it lyrics tagged [Verse]/[Chorus] plus a caption for genre, BPM, key, and instrumentation.
It runs on your own GPU
Full precision wants ~24GB VRAM, CPU offload trims it to ~22GB, and streaming mode drops to 8GB — so a single 24GB card, or even an 8GB one, is enough. Day-one GGUF quants plus ComfyUI and Diffusers support mean you can wire it in tonight.
The license lets you ship
The MiniMax-Music3 Community License reads MIT-style (“free… to deal in the Software”) and clears commercial use unless your company tops $20M in annual revenue. Unlike its H3 model, there is no US/EU/UK/Korea carve-out this time.
Qwen-backed, caption-controlled
An 8B global LLM initialized from Qwen3.5-8B handles song structure while a 0.6B local LLM and a flow-matching synth stack fill in the audio. New “structured captions” steer style section by section — the big quality jump over Music 2.
Elsewhere: Google open-sources an FHE compiler
Google’s HEIR toolchain (Aug 14) compiles a pre-trained model to run inference on fully encrypted data, aiming at one-click encrypted inference for non-experts. Demo apps — fraud detection, hotword spotting — ship on GitHub.