MiniMax Music 3 — Local AI Song Generation
MiniMax Music 3 is a new open weights AI music generation model, and it's one of the strongest local alternatives to Suno yet. Give it a set of lyrics and a description of the sound you want, and it generates a complete, structurally coherent song — vocals, instruments, and arrangement — up to five minutes long in 32 kHz stereo. It runs natively in ComfyUI under MiniMax's own community license, so there's no subscription, no per-song credit meter, and nothing leaves your machine.
✓ 100% free — download the workflow and run it locally.Requirements
- NVIDIA RTX 30XX / 40XX / 50XX GPU (also runs on Ada-generation workstation cards)
- CUDA-compatible GPU (~8 GB VRAM should be enough for the FP16 build, though I haven't fully tested the low end)
- At least 20 GB free storage
Model downloads
- minimax_music3_dit_int8_convrot.safetensors – low VRAM / smaller option https://huggingface.co/Comfy-Org/MiniMax-Music-3/tree/main/diffusion_models
- minimax_music3_text_encoder_pruned_int8_convrot.safetensors https://huggingface.co/Comfy-Org/MiniMax-Music-3/tree/main/text_encoders
- minimax_music3_dav.safetensors https://huggingface.co/Comfy-Org/MiniMax-Music-3/tree/main/vae
Custom nodes
Setup steps
- Run in the Cloud, and More Local AI Video
- Full songs, not clips — complete tracks up to 5 minutes with real structure (intro, verses, chorus, outro), not 10-second loops.
- Expressive vocals with your lyrics — it sings the words you give it, holding a consistent vocal identity across the song.
- Runs local and light — small models, ~8 GB VRAM territory, and completely free to run at home.
- Simple workflow — two text inputs (a caption and your lyrics) feed one node that outputs a finished song.
- How to Write the Caption (This Matters More Than Anything)
- Global metadata — genre, BPM, key/scale, emotional progression, and production profile.
- Vocal details — vocal gender, tone, and delivery style.
- Arrangement — how the sections evolve across the song.
Notes & tips
- Load the provided workflow into ComfyUI.
- Verify the three models are selected in the node dropdowns — the FP16 DiT (in unet), the text encoder (in clip), and the DAV VAE (in vae). If any dropdown is blank, the file is in the wrong folder.
- Write your caption in the three-section structure above, and tag your lyrics with [Verse] / [Chorus] on their own lines.
- Keep your first generations short to dial in the caption and vocal style before committing to a full 5-minute track.
- Generation time scales with song length and depends heavily on your GPU. On a lower-end card it can take a while for a full-length song, so a stronger GPU makes a big difference if you're iterating.
- Use the FP16 DiT for final renders — the quality gain over INT8 is worth it if you have the VRAM.
More resources
Want the 1-click version?
Skip the manual model downloads and node installs — get a ready-to-run installer, or unlock all 75+ with Local Lab Pro.