Skip to main content

MiniMax's Music3: Turn Lyrics into Full Songs in Minutes

What if you could hum a tune, jot down some lyrics, and have a full song ready in minutes? That's the promise of MiniMax's latest creation, the Music3 music generation model. Launched recently, this AI tool takes your words and a simple description, then spins them into a complete track that can run up to five minutes long—no instruments, no studio, no problem.

The output is a 32kHz, 16-bit stereo WAV file, which sounds pretty impressive for something generated from scratch. But the real magic lies in how it works under the hood.

A Tale of Two Models

Music3 isn't just one big neural network doing everything. Instead, it's a clever duo, each with its own job. Think of it like a band: one member handles the big picture, while another fills in the details.

The first is the Global LLM, a hefty 8-billion-parameter model that predicts the song's overall structure—the verses, the chorus, the bridge, all that. It's like the composer who sketches out the entire piece, deciding when the energy rises and falls. This model is actually initialized from Qwen3-8B, a well-known language model, but it's been adapted to think in musical terms.

The second is the Local LLM, a much smaller 0.6-billion-parameter model. Its job is to predict the remaining acoustic details, the fine-grained sounds that make a song feel alive—the subtle guitar strum, the breath in a vocalist's voice, the cymbal crash. If the Global LLM is the architect, the Local LLM is the interior designer, making sure every corner feels right.

Together, they work in a hierarchical autoregressive fashion. The Global LLM predicts the first codebook frame by frame, setting the semantic and structural foundation. Then the Local LLM fills in the rest, layer by layer, until the song is complete.

Keeping It Together

One of the biggest challenges in AI music generation is maintaining coherence over long pieces. Many models start strong but lose the thread halfway through, resulting in a mess of disconnected sounds. MiniMax claims Music3 handles this beautifully, keeping the musical theme, rhythm, vocal identity, and arrangement progression steady from start to finish.

That means you get a song that actually feels like a song—with a proper intro, verses, a catchy chorus, maybe a bridge, and a satisfying outro. No more random noise or abrupt shifts. The official announcement even highlights that the model can hold onto these elements across the entire five-minute duration, which is no small feat.

Try It Yourself

If you're curious to hear what Music3 can do, the team has already released the model page on GitHub, complete with sample songs for you to listen to. You can judge for yourself whether it lives up to the hype.

This launch is part of a broader trend in AI music generation, with companies like Suno and Udio also pushing the boundaries. But MiniMax's approach—using a large model for structure and a small one for details—seems particularly efficient. It's a smart way to balance quality and computational cost.

What This Means for Music

For musicians, this could be a game-changer. Imagine being able to quickly generate a demo track to share with a bandmate, or experimenting with different arrangements without spending hours in a studio. For hobbyists, it's a fun way to bring their lyrical ideas to life, even if they can't play an instrument.

Of course, there are questions about copyright and originality, but that's a conversation for another day. For now, Music3 is a fascinating step forward in making music creation more accessible.

So, if you've ever had a song stuck in your head but couldn't get it out, maybe this is your chance. Write down those lyrics, describe the vibe you're going for, and let Music3 do the rest. Who knows? You might just create the next big hit—or at least a catchy tune for your next road trip.

Key Points

  • MiniMax-Music3 generates complete songs up to 5 minutes long from lyrics and descriptions.
  • Dual-model architecture: an 8B Global LLM handles structure, a 0.6B Local LLM fills in acoustic details.
  • Output: 32kHz, 16-bit stereo WAV files.
  • Coherence: maintains theme, rhythm, and arrangement across long tracks.
  • Availability: model page and samples on GitHub.