Skip to main content

MiniMax's New Music3 Model Crafts 5-Minute Songs from Simple Prompts

Imagine typing a few lines of lyrics, adding a note like "upbeat pop with a hint of nostalgia," and getting back a complete, polished song—intro, verses, chorus, bridge, the works. That's the promise of MiniMax's latest creation, the Music3 music generation model. Launched recently, this AI doesn't just string together notes; it crafts a full musical journey that can stretch up to five minutes long, delivered as a high-quality 32kHz, 16-bit stereo WAV file.

What makes Music3 stand out in the crowded field of AI music tools? It's all about the architecture. The team behind it has designed a hierarchical system where two distinct AI models work in tandem, each with a clearly defined role. Think of it as a songwriting duo: one handles the big picture, the other obsesses over the tiny details.

The first model, dubbed the Global LLM, is the big one—8 billion parameters, to be precise. It's initialized from Qwen3-8B, a well-known language model, but fine-tuned for music. Its job is to predict the first RVQ codebook frame by frame, which essentially means it maps out the song's long-term structure and semantic flow. In other words, it decides how the song unfolds, ensuring the theme, rhythm, and vocal identity stay consistent throughout. It's the architect of the track.

Then there's the Local LLM, a more compact model with just 0.6 billion parameters. Despite its smaller size, its role is just as crucial: it predicts the remaining acoustic codebooks for each frame, filling in the fine-grained audio details that make the song sound rich and alive. If the Global LLM is the architect, the Local LLM is the interior designer, adding the textures and nuances that turn a skeleton into a living, breathing piece of music.

This division of labor is what allows Music3 to maintain a firm grip on the musical theme, rhythm, and arrangement progression over long durations. The official release notes boast that the model doesn't miss any structural elements—prelude, verse, pre-chorus, chorus, bridge, instrumental interlude, and coda all come together seamlessly. For anyone who's tried other AI music generators, this is a big deal. Often, these tools struggle to keep a song coherent beyond a minute or two, but Music3 aims to change that.

Image

So, how does it actually work in practice? You provide lyrics and a description of the musical style you want. The model then generates a complete song, complete with vocals and instrumentation, based on that input. The output is a WAV file, which is a lossless format, meaning you get studio-quality audio right out of the box.

MiniMax has already made the model page available on GitHub, where you can also listen to sample songs. It's a smart move—letting the music speak for itself. And for developers and music enthusiasts, this opens up a world of possibilities. Imagine being able to prototype a song idea in minutes, or create custom background music for videos without needing a full production setup.

But let's step back for a moment. Why should you care? If you're a musician, this could be a powerful tool for inspiration or even for generating demos. If you're a content creator, it's a way to get original music that fits your needs without licensing headaches. And if you're just someone who loves music, it's fascinating to see how AI is pushing the boundaries of creativity.

Of course, there are questions about the quality and originality of AI-generated music. Can a machine truly capture the emotional depth of a human-composed piece? The samples on GitHub suggest that Music3 is getting closer, but it's still early days. The model's ability to maintain structure over five minutes is impressive, but the soul of the music—that indefinable something—remains to be seen.

One thing's for sure: the team at MiniMax is betting big on this technology. By combining a large model for structure and a small one for details, they've found an efficient way to generate long-form audio without sacrificing quality. It's a clever hack that could inspire other AI music projects.

So, whether you're a tech enthusiast, a musician, or just curious about where AI is headed, Music3 is worth a listen. Head over to GitHub, check out the samples, and maybe even try it out yourself. Who knows? Your next favorite song might just be written by an AI.

Key Points

  • Music3 generates songs up to 5 minutes long from lyrics and descriptions, outputting 32kHz stereo WAV files.
  • Two-model architecture: an 8B-parameter Global LLM handles song structure, while a 0.6B-parameter Local LLM fills in acoustic details.
  • Maintains musical coherence across long tracks, including all standard song sections.
  • Available on GitHub with sample songs for listening.
  • Potential applications include music prototyping, content creation, and creative inspiration.