Skip to main content

MiniMax Music3: Turn Lyrics into Full-Length Songs in Minutes

MiniMax has just dropped a new AI model that might change how you think about music creation. Called Music3, it can take a simple set of lyrics and a brief description of the vibe you're going for, and turn them into a full song—up to five minutes long. No instruments, no studio, just a few words and a whole lot of AI magic.

The output is a 32kHz, 16-bit stereo WAV file, which sounds pretty professional for something generated from scratch. But what's really interesting is the tech underneath.

A Tale of Two Models

Music3 isn't just one big neural network doing everything. Instead, it's a duo—two models with distinct jobs, working in harmony. Think of it like a band where one member handles the big picture while another fills in the details.

The first is a global language model with 8 billion parameters. It's the architect, mapping out the song's long-term structure and semantics. It predicts the first codebook frame by frame, essentially deciding how the song flows from start to finish. Interestingly, it starts from Qwen3-8B, a model originally designed for text, but it's been adapted to think in musical terms.

The second model is much smaller—just 0.6 billion parameters—but it's no less important. This local model takes over the remaining acoustic codebooks in each frame, layering in the fine-grained audio details that make the song sound rich and alive.

So, one big brain handles the skeleton, and a nimble sidekick fills in the flesh. It's a division of labor that seems to work remarkably well.

Keeping It Together

One of the biggest challenges in AI music generation is maintaining coherence over long stretches. Many models start strong but lose the thread after a minute or so. Music3, according to MiniMax, holds up impressively. It keeps the musical theme, rhythm, vocal identity, and arrangement progression intact throughout the entire track.

And it doesn't skip any structural elements either. You get the prelude, verse, pre-chorus, chorus, bridge, instrumental interlude, and coda—all the pieces you'd expect from a well-crafted song.

Image

Try It Yourself

MiniMax has already released the model page on GitHub, complete with sample songs you can listen to. So if you're curious about what this AI can do, you don't have to take their word for it—give it a listen and see if it hits the right notes.

Key Points

  • Music3 generates complete songs up to 5 minutes from lyrics and descriptions, outputting 32kHz stereo WAV.
  • Two-model architecture: an 8B global LLM handles structure, while a 0.6B local LLM fills in acoustic details.
  • Long-form coherence: maintains theme, rhythm, vocal identity, and arrangement throughout.
  • Full song structure: includes prelude, verse, chorus, bridge, and more.
  • Available on GitHub with sample tracks for listening.