MiniMax · 2026-08-13 · major
MiniMax Music 3.0 — open-weights model writes a full five-minute song
MiniMax Music 3.0 turns a short concept and optional lyrics into a finished song of up to five minutes in one pass. MiniMax published the weights on Hugging Face under CC-BY-SA 4.0 with inference code on GitHub.
An open-weights music model that composes, arranges, performs and produces a whole song in a single pass.
Quick facts
| Maker | MiniMax |
|---|---|
| License | CC-BY-SA 4.0 |
| Max song length | 5 minutes |
| Audio output | 32 kHz, 16-bit stereo WAV |
| Language models | 8B global + 0.6B local |
| Synthesis stack | 2.4B flow matching + 123M Flow-VAE |
| Hardware | Two CUDA GPUs |
What is it?
MiniMax Music 3.0 generates a complete song — composition, arrangement, vocals and production — from a creative concept and optional lyrics, without stitching separate stages together. Tracks can run up to five minutes and come back as 32 kHz, 16-bit stereo WAV. MiniMax released the weights publicly alongside the inference code, so the model can be run locally rather than only through an API.
How does it work?
A hierarchical autoregressive design splits the work between two language models. The 8B global model predicts the first residual-vector-quantization codebook frame by frame and carries the song's long-range structure; a 0.6B local model fills in the remaining acoustic codebooks inside each frame. The tokenizer uses eight RVQ layers — one 16,384-entry codebook for core semantics plus seven 1,024-entry codebooks for acoustic detail. A 2.4B flow-matching module then maps the fused hidden states into VAE latent space, and a 123M Flow-VAE decoder reconstructs the waveform.
Why does it matter?
Published weights change who can build on a music model of this class. Studios and tool makers can host MiniMax Music 3.0 themselves, inspect the architecture and adapt it, instead of renting a closed endpoint and accepting whatever it returns. The practical limit is hardware — two CUDA GPUs makes this a server or workstation model, not something that runs on a laptop.
Who is it for?
audio ML engineers and music tool builders
Frequently asked questions
- Can you use MiniMax Music 3.0 commercially?
- MiniMax Music 3.0 ships its weights under CC-BY-SA 4.0. That license permits redistribution and adaptation, including for commercial work, as long as you credit MiniMax and release anything derived from the model under the same share-alike terms. Check the LICENSE file in the repository before shipping a product on top of it.
- What hardware do you need to run MiniMax Music 3.0?
- Running MiniMax Music 3.0 takes two CUDA GPUs. The first handles autoregressive token generation and the second runs synthesis and waveform decoding. MiniMax serves the model through SGLang-Omni, and the current release supports non-streaming generation only, so a song is returned as one finished file rather than progressively.
- How do you control what MiniMax Music 3.0 writes?
- Two inputs steer MiniMax Music 3.0. Lyrics go in the request's input field and can carry section tags such as Verse, Chorus and Bridge to shape the song structure. A separate instructions field takes a written music description covering genre, mood, instruments and vocal style, so you can specify an arrangement without supplying any lyrics at all.
- Where can you download MiniMax Music 3.0?
- The MiniMax Music 3.0 weights are on Hugging Face as MiniMaxAI/MiniMax-Music3, and the inference code, tokenizer details and serving instructions live in the MiniMax-AI/MiniMax-Music3 repository on GitHub. The repository also includes a music-caption-rewriter skill that expands a short description into the structured caption format the model expects.
Try it
sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000