Official facts · 13 August 2026
What MiniMax Music 3.0 is
MiniMax Music 3.0 is a music generation model. Give it a creative idea and optional lyrics; in one generation it composes, arranges, sings, and produces a complete song. That definition is MiniMax’s, published on 13 August 2026.
Source: English research post · 中文官方博文 · Hugging Face weights
What is MiniMax Music 3.0?
MiniMax Music 3.0 is the company’s next-generation open-weight music model. It is built for the parts of songwriting that a one-line prompt usually misses: holding a creative intention across a full song, rendering instruments with playing-technique texture, and making vocals sound sung rather than synthesized. Official documentation states that a single pass can finish composition, arrangement, vocal performance, and production, with complete songs up to five minutes.
How the model is built
MiniMax describes three linked stages. An eight-layer residual vector quantizer separates musical structure from acoustic residue. A Hybrid-LM then models the song: an 8B global language model, initialized from Qwen3.5-8B, predicts semantic tokens frame by frame, while a 0.6B local model predicts in-frame acoustic tokens. Synthesis fuses the two models’ continuous hidden states into a 2.4B flow-matching module and a 123M Flow-VAE decoder. The published chain is fused LLM features → flow-matching → VAE latents → stereo audio.
What you can run today
- Official web creator on MiniMax Audio (Music-3.0, limited free beta)
- Official API IDs
music-3.0andmusic-3.0-free - Open weights on Hugging Face as MiniMaxAI/MiniMax-Music3
- Local paths: ComfyUI, Diffusers, SGLang-Omni (CUDA)
Create a track in the studio, or read the guide.
Last reviewed 2026-08-14 against the official 13 Aug 2026 posts.