Introduction

Feel that? The room just got heavier. No sooner then the controversy with Suno's new limitations, a new open-weight model just dropped and it’s swinging with the kind of momentum that makes every other generator feel like a light sketch. MiniMax Music 3 doesn’t spit out short loops or half-formed ideas. It delivers complete songs—up to five full minutes—complete with structure, expressive vocals, and arrangements that actually evolve. One detailed music description plus tagged lyrics, and the track rolls out with the unfiltered motion of a live performance.

This is the model that finally treats long-form music like a real composition instead of a lucky accident.

The Architecture Behind the Heavy Impact

MiniMax Music 3 is built as a hybrid system that separates the big picture from the fine grain. An 8B Global LLM (initialized from Qwen3-8B) handles long-range structure—themes, rhythm, vocal identity, and how the arrangement develops across an entire song. A much smaller 0.6B Local LLM fills in the frame-level acoustic details so every note and breath still feels alive.

Those continuous hidden states then feed a Flow Matching stage and a Flow-VAE decoder, producing clean 32 kHz, 16-bit stereo WAV audio. The result is music that keeps its shape from intro to outro without collapsing into the usual AI mud.

Pro Tip: The model responds best to a Structured Caption. Break it into Global Metadata (genre, BPM, key, emotional arc), Vocal Details (gender, timbre, performance style, harmonies), and Arrangement (instruments, section-by-section evolution, groove, spatial effects). The difference in coherence is night and day.

What Makes It Swing

  • Full songs up to five minutes with native long-range coherence
  • Explicit section tags in lyrics: [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Outro], and more
  • Expressive, controllable vocals that stay consistent across the whole track
  • Rich genre coverage—from rock, R&B, and hip-hop to blues, bossa nova, metal, Chinese regional styles, and Japanese vocals
  • Open weights under the MiniMax-Music3 Community License

You can hear the range for yourself on the official demo page: https://minimax-ai.github.io/music3-demo/. Dozens of full tracks across styles prove the model isn’t just competent—it moves.

How to Get Started

The weights live here: https://huggingface.co/MiniMaxAI/MiniMax-Music3

Want to try it without downloading anything first? Jump straight into the interactive Space: https://huggingface.co/spaces/MiniMaxAI/MiniMax-Music3

Local Inference Options

  1. Download the model:

hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm

  1. Serve it with SGLang-Omni (recommended for full control):

sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000

  1. Generate with a simple API call. Put lyrics (with section tags) in input and the music description in instructions:

curl http://127.0.0.1:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "minimax_ttm",
    "input": "[Verse]\nMorning light filtering through the pine\n[Chorus]\nSoftly the world begins to breathe",
    "instructions": "A warm acoustic pop song with intimate female vocals, fingerpicked guitar, soft piano, and a gradual emotional build into a wide final chorus.",
    "response_format": "wav",
    "seed": 7,
    "max_new_tokens": 9000,
    "stream": false
  }' \
  --output minimax_music3.wav

ComfyUI users get a ready-made workflow in recent versions—just load the official template, drop in your caption and lyrics, and let it run. Diffusers support is also landing for those who prefer the modular pipeline route.

There’s even a handy skill that expands a short idea into a full Structured Caption:

npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter

Conclusion

MiniMax Music 3 doesn’t just generate audio—it carries weight. The deep low-end stays locked, the arrangements keep moving, and the vocals hold their character from the first bar to the final fade. Whether you’re writing for games, films, personal projects, or just chasing that next unforgettable hook, this open-weight model finally gives creators the long-form control we’ve been waiting for.

Grab the weights, spin up the Space, or dive into the demos. The rhythm is already there—now go make it swing.

—Anabel ♡