MiniMax Music 3 makes 5-minute songs on your own GPU, with no subscription or download cap

MiniMax put the weights for its new song generator on Hugging Face on August 13, 2026. Feed MiniMax Music 3 lyrics and a written description of the sound you want, and it returns a finished song of up to 5 minutes at 32kHz, 16-bit stereo, with vocals and a full arrangement.

The download is the product. There is no account, no monthly quota, and no server deciding when your render finishes.

How MiniMax Music 3 keeps a 5-minute song from falling apart

Long songs break because a model forgets what it played 2 minutes ago. MiniMax splits the job across two models running together. An 8B Global LLM tracks the song’s structure and long-range direction frame by frame, while a 0.6B Local LLM fills in the fine acoustic detail inside each frame.

The Global model started life as Qwen3-8B, a text model, and was retrained on music tokens. Synthesis runs on Flow Matching and a Flow-VAE adapted from MiniMax’s speech work, which is where the audio quality comes from.

Control comes through two inputs. Lyrics carry section tags like [Verse], [Pre-Chorus], [Bridge] and [Outro], and a separate music description sets genre, BPM, key, vocal performance, instrumentation and production feel. MiniMax ships a caption-rewriter skill that expands a short idea into the structured prompt the model prefers.

The honest limits are on the model card. Inference needs CUDA, so Mac users are out until someone ports it. Generation is non-streaming, prompts cap at 5,000 tokens, and MiniMax says tempo, key and structure requests are guidance rather than a guarantee.

What the MiniMax-Music3 Community License lets you release

The license is close to MIT with three additions worth reading before you upload anything. You have to show “MiniMax-Music3” on the interface of any commercial product built on the model. Past $20 million in yearly revenue you need written permission from MiniMax. And if you let other people generate through your service, you have to build and maintain safeguards against infringing output.

For a producer putting a track on Spotify, none of those bite. For anyone building a tool on top, all three do.

The gap sits where it always sits. MiniMax documents its architecture in detail and names its code lineage, including Qwen3-8B under Apache 2.0 and pieces adapted from Stable Audio and Descript’s audio codec. It names no music. A German court ruled against Suno in July on exactly that question, and running a model on your own hardware moves the copyright risk to you rather than removing it.

What open weights do change is the economics. Suno starts capping downloads on September 3, with 7 lifetime downloads on Free and 20 a month on Pro. A local model has no such ceiling, and it is the first full-song generator good enough to make that tradeoff a real one.