AEmotionStudio/minimax-music-3-models

MiniMax Music 3 — MAESTRO mirror

Redistribution of the diffusers-layout subset of MiniMaxAI/MiniMax-Music3 (MiniMax, August 2026), packaged for the MAESTRO audio workstation's offline runtime. All credit to MiniMax. Weights are released under the MiniMax-Music3 Community License (reproduced verbatim at shared/LICENSE): commercial use is permitted provided the product prominently displays "MiniMax-Music3"; services above US$20M/year revenue need separate authorization from MiniMax. This mirror adds no restrictions and claims no ownership. MAESTRO's inference code for this model is vendored separately under Apache-2.0.

MiniMax Music 3 generates full songs (lyrics + style prompt) up to five minutes: an 8B language model plans the codec sequence, a diffusion transformer (DiT) renders it, an RVQ depth decoder + vocoder produce the waveform.

Layout

shared/    modular_model_index.json, config.json, LICENSE
           tokenizer/**, scheduler/**, condition_encoder/**, vocoder/**,
           rvq_depth_decoder/** (bf16, verbatim),
           transformer/** (the DiT — upstream fp32, CAST to bf16 here, see below)
quality/   language_model/** — the 8B LM, four bf16 shards VERBATIM upstream

shared/ + quality/ (22.0 GB) is exactly what the app's "Quality" manifest downloads. A fast/ lane (house int8 W8A8 quantization of the language model) is described by the app's second manifest but is not present in this mirror: the int8 build was measured to collapse on long autoregressive runs and was not published.

What was changed

• shared/transformer/ (the DiT) ships fp32 upstream (9.7 GB); MAESTRO loads it in bf16 regardless, so it is cast once here (~4.9 GB, no runtime difference). Every other tensor is byte-for-byte upstream.
• The SGLang-only payload (qwen_7B/, flowmatching_vae.pth, dav.pth) and the upstream remote-code files are not mirrored — MAESTRO vendors its own diffusers-style pipeline and never uses trust_remote_code.

Loading

Inside MAESTRO: backend/ai/models/minimax_music_3.py (_resolve_paths reads this shared/ + quality/ layout; the language model is placed with the shared offload planner on 12 GB cards). Users can also point the runner at the upstream repo directly.

License

MiniMax-Music3 Community License — © 2026 MiniMax. See shared/LICENSE. Please credit MiniMax and display "MiniMax-Music3" as the license requires.