MiniMax Music 3

You can download this one and run it. MiniMax Music 3 is a 57GB repository on Hugging Face pairing an 8B global model for song structure with a 0.6B local model for frame-level detail, and it will write a five-minute song with vocals at 32kHz stereo on hardware you own. ComfyUI shipped a template workflow on day one and community GGUF quantizations followed within days. All of that is real, and it landed in the same fortnight as the Suno caps that took effect today. Then read the license. The weights ship under MiniMax's own community terms, which are not an OSI license and do not let you sell what comes out, and no benchmark numbers are published against Suno, Udio, YuE or ACE-Step. Open enough to inspect, not open enough to build on. Publish the benchmarks, or change the license, and this becomes a different entry.