MiniMax-Music3: self-hostable AI for full 5-minute songs
Open weights, a permissive commercial license, and it runs down to 8GB VRAM — plus DuckDB's v2.0 preview adds similarity-search joins for RAG.
Copy markdown
Full songs from lyrics and a caption
MiniMax-Music3 is open-weights and generates complete tracks up to five minutes — vocals and arrangement — from tagged lyrics ([Verse], [Chorus]) plus a structured caption, output as 32 kHz stereo WAV. Weights, inference code, and three serving paths shipped together.
This time, no territorial carve-out
The Community License grants free commercial use with attribution — unlike MiniMax's H3 video model days earlier, which excluded the US, EU, UK and South Korea from local deployment. Only companies above $20M/year in revenue need written sign-off, so indie builders are in the clear.
Runs on the GPU you already have
Hugging Face Diffusers group-offloads down to ~8GB VRAM (24GB at full precision), and a native ComfyUI template ships FP16/INT8 weights on day one. That's commercially usable background music and game or video soundtracks you can generate locally this weekend.
Elsewhere: DuckDB v2.0 preview lands similarity-search joins
The v2.0 preview (stable this fall) adds NEAREST joins for similarity search, a client/server "quack" mode, and a VARIANT semi-structured type — useful if you run local RAG or analytics right next to your models.