MiniMax-Music3
#6 en Génération musicale IAMinimax · v3 · 2× · vu le 17 août 2026
MiniMax-Music3 (also referred to as "Music 3.0") is an AI music generation model developed by MiniMax that generates complete songs of up to five minutes in length from song lyrics and music descriptions, including structure (intro, verse, chorus, bridge, etc.), vocals, and arrangement. It is based on a hierarchical architecture featuring an 8B-Parameter Global LLM for long-term musical structure and a 0.6B-Parameter Local LLM for acoustic details, combined with a Flow-Matching and Flow-VAE synthesis system. The model weights are openly available under a proprietary Community License (Hugging Face, GitHub, ModelScope), with a hosted API variant also available. Output is delivered as 32-kHz 16-bit stereo audio.
Fonctionnalités
| Real-Time Streaming | Local inference currently supports only non-streaming generation; hosted API offers optional streaming (hex output only) |
| License | MiniMax-Music3 Community License (commercial use allowed, separate written authorization required above $20M annual revenue, mandatory attribution) |
| Platform | Hosted Music Generation API and downloadable weights (Hugging Face, GitHub, ModelScope) for local inference via SGLang-Omni, diffusers, or ComfyUI |
| Price | $0.15 per hosted API generation; free music-3.0-free tier available; open weights free under Community License |
| Release Date | API: July 16, 2026; open weights: August 13, 2026 |
| Languages | Supports lyrics in many languages |