Model · MiniMax · Audio-Gen
MiniMax Music 3
MiniMax's open-weight song generator. Give it lyrics and a description and it composes, arranges and sings a complete song up to five minutes long, as 32 kHz stereo audio.
- Modality
- Audio-Gen
- License
- MiniMax-Music3 Community License (Open weights)
- Parameter size
- 8B global LLM + 0.6B local LLM + 2.4B flow-matching decoder
- Released
- August 13, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
What you need to run it
- Mac
- No. Needs an NVIDIA graphics card — there is no working Mac version.
- Windows / Linux PC
- NVIDIA only. A graphics card with at least 8 GB (24 GB recommended) of video memory. AMD and Intel cards are not supported.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
The official code requires an NVIDIA card; a community MLX port exists for Macs (the 4-bit version is about 8 GB on disk) but it is experimental and not made by MiniMax.
Numbers are for: Full-precision weights (fit 24 GB; 8 GB with slow layer-by-layer offloading). Based on the maker’s own figures; where the maker gives only a graphics-card figure, the Mac memory is our estimate.
Strengths
- Complete songs with vocals up to five minutes, with coherent structure (verse, chorus, bridge)
- Fine control through lyrics with section tags plus a structured music description (genre, BPM, key, vocals, instruments)
- Full precision fits under 24 GB of graphics memory; with layer-by-layer offloading it fits an 8 GB card, more slowly
- Runs in SGLang-Omni, diffusers and ComfyUI
Weaknesses
- Commercial use requires showing "MiniMax-Music3" prominently in your product's interface (MiniMax-Music3 Community License)
- Businesses with over US$20 million yearly revenue must get written permission from MiniMax before commercial use
- Low-memory mode is markedly slower than running on a 24 GB card
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | MiniMax-Music3 Community License |
| GitHub | local | Code and caption-rewriter skill |
| ComfyUI | local | Node workflow tutorial |
Official sources
- Model card model
- Licence docs
- Announcement announcement
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: the model card, which states 'Inference requires CUDA'.
- — Initial entry. Added 2026-10-28 from the popularity sweep: 1,484 Hugging Face likes, 992 GitHub stars. Released 2026-08-13 per the MiniMax blog post. Licence terms read from the LICENSE file (last revised August 2026): commercial use allowed if 'MiniMax-Music3' is displayed in the product UI, with written permission needed above US$20M yearly revenue. Global LLM is fine-tuned from Qwen3-8B.
Esc