Operator notes

How to run MiniMax Music 3.0

There are four real doors. Use the one that matches the job, not the hype. Architecture facts below follow MiniMax’s 13 Aug 2026 Music 3.0 post.

1. Official web creator

Fastest listen. MiniMax Audio currently lists Music-3.0 with a limited-time free creator beta and optional song-reference cover. Start at minimax.io/audio/music.

2. Official API

Endpoint: POST https://api.minimax.io/v1/music_generation. Model IDs that matter this week:

• music-3.0 — paid / token plan, RPM 120
• music-3.0-free — API key, RPM 3
• music-cover / music-cover-free — reference cover

Send a prompt (style, mood, scene) and lyrics with section tags, or set is_instrumental: true, or lyrics_optimizer: true to write lyrics from the prompt. Docs: music generation guide.

curl https://api.minimax.io/v1/music_generation \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "music-3.0-free",
    "prompt": "Indie folk, 92 BPM, D major, intimate female vocal",
    "lyrics": "[Verse]\nRain writes your name across the glass\n[Chorus]\nIf the night is a borrowed coat",
    "output_format": "url"
  }'

3. Local open weights

Weights: MiniMaxAI/MiniMax-Music3. Architecture in the model card: 8B global LLM + 0.6B local LLM + flow-matching synthesis → 32 kHz 16-bit stereo WAV. Serve with SGLang-Omni or Diffusers. CUDA required. Prompt token cap 5,000. Audio frame cap 9,000.

4. ComfyUI

Day-one Comfy support shipped with the weights. Follow the official ComfyUI Music 3 tutorial or our shorter ComfyUI page.

Structure tags Music 3 understands

Put tags on their own lines: [Intro] [Verse] [Pre-Chorus] [Chorus] [Post-Chorus] [Bridge] [Instrumental] [Solo] [Outro] (API also lists Hook, Break, Build Up, Inst).

For better control, write the description as three blocks: Global Metadata, Vocal Details, Arrangement. That is the format MiniMax recommends and the format this studio emits.