Skip to main content

MiniMax Music3: Turn Lyrics into Full Songs Up to 5 Minutes

Imagine typing in a few lines of lyrics, adding a note like "upbeat pop with a catchy chorus," and getting back a fully produced song that lasts five minutes. That's exactly what MiniMax's new Music3 model promises. The company recently launched this music generation tool, and it's already turning heads in the AI community.

Music3 works by taking two inputs: your lyrics and a description of the musical style you want. From there, it generates a complete track, complete with vocals, instrumentation, and even structure—think intro, verse, chorus, bridge, and outro. The output is a 32kHz, 16-bit stereo WAV file, which is pretty high quality for an AI-generated tune.

What's really interesting is how the model is built. It uses a hierarchical autoregressive architecture, which sounds complicated, but the idea is simple: two models team up, each with a specific job. The first is a large "global" language model with 8 billion parameters. It's responsible for the big picture—the song's overall structure and long-term musical themes. It predicts the first codebook frame by frame, essentially mapping out the song's skeleton. This model is initialized from Qwen3-8B, a well-known language model, and during training, it's adapted to understand musical semantics before teaming up with the local model.

The second is a smaller "local" model, with just 0.6 billion parameters. Its job is to fill in the details—the fine-grained acoustic information that makes the song sound rich and complete. It predicts the remaining codebooks in each frame, adding layers of sound that bring the skeleton to life.

This division of labor is pretty clever. One big model handles the structure, and a small one handles the details. It's like having an architect and an interior designer working on the same house. The architect draws the blueprint, and the designer picks the paint colors and furniture.

According to MiniMax, this approach allows the model to maintain the musical theme, rhythm, and vocal identity throughout long audio pieces. It won't lose track of the song's flow or forget to include a bridge or an instrumental break. The official announcement highlights that the model can handle all the typical song sections—prelude, verse, pre-chorus, chorus, bridge, instrumental interlude, and coda—without missing a beat.

If you're curious to hear what Music3 can do, you're in luck. The model page is already live on GitHub, and the team has posted sample songs for you to listen to. So you can judge for yourself whether AI-generated music is ready for the radio.

This launch is part of a broader trend in AI music generation. Companies like OpenAI and Google have been working on similar tools, but MiniMax's focus on long-form songs (up to five minutes) sets it apart. Most other models struggle to maintain coherence beyond a minute or two, so this is a significant step forward.

Of course, there are still questions about copyright and originality. If an AI writes a song based on your lyrics, who owns the rights? And can AI truly capture the emotional nuance that human composers bring? These are debates that will likely continue as the technology evolves.

For now, Music3 is a fun and powerful tool for musicians, hobbyists, and anyone who's ever hummed a tune and wished they could turn it into a full track. It's not going to replace human creativity, but it might just be the spark you need to get started.

So, whether you're a seasoned producer or just someone who likes to play around with new tech, give Music3 a try. Who knows? Your next favorite song might be written by an AI—with a little help from you.

Key Points

  • MiniMax Music3 generates complete songs up to 5 minutes long from lyrics and a description.
  • Output format: 32kHz, 16-bit stereo WAV file.
  • Architecture: A hierarchical autoregressive model with two components—a global LLM (8B parameters) for structure and a local LLM (0.6B parameters) for acoustic details.
  • Capabilities: Maintains musical theme, rhythm, and vocal identity across long tracks, including all standard song sections.
  • Availability: Model page and sample songs are on GitHub.