MiniMax-Music-3 — GGUF
GGUF quantizations of MiniMax-Music3, converted from Comfy-Org's fp32 repack for ComfyUI with ComfyUI-GGUF.
MiniMax-Music-3 generates full music from a text prompt. The diffusion transformer is only 2.46B parameters — small enough that even the top tier fits comfortably on modest hardware, and Q4 runs on almost anything.
Apache-2.0, inherited from the base model.
Files
| Quant | Size | Notes |
|---|---|---|
| Q6_K | ~2.0 GB | Excellent |
| Q5_K_M | ~1.8 GB | Very good |
| Q4_K_M | ~1.5 GB | Good balance |
| Q3_K_M | ~1.2 GB | Audible softening on transients |
| Q2_K | ~0.9 GB | Smallest; noticeable artifacts |
hf download realrebelai/MiniMax-Music-3-GGUF MiniMax-Music-3-Q8_0.gguf --local-dir ComfyUI/models/diffusion_models
Take a higher tier than you would on an image model. The fp32 source is 9.15 GB and Q8_0 is 2.6 GB — on a model this small there's little reason to drop low, and audio gives quantization error nowhere to hide. There's no visual channel to mask it: what would read as slightly soft detail in an image reads as smeared transients, ringing, or a thinner stereo field here. Q8_0 or Q6_K is the sensible default.
Companion files
The GGUF replaces only the diffusion transformer. Both of these come from Comfy-Org/MiniMax-Music-3:
| File | Folder |
|---|---|
minimax_music3_text_encoder_bf16.safetensors |
models/text_encoders |
minimax_music3_dav.safetensors |
models/vae |
A pruned text-encoder variant is also available in that repo if the bf16 encoder is too large for your setup.
Usage
- Put the
.ggufinComfyUI/models/diffusion_models - Load it with Unet Loader (GGUF)
- Text encoder, VAE and sampler settings are unchanged from the official MiniMax-Music-3 workflow
Requires ComfyUI with MiniMax-Music-3 support, plus the ComfyUI-GGUF custom node.
Conversion notes
Converted from the official fp32 transformer with a streaming converter, quantized with llama-quantize, then repaired.
This is a 1-D convolutional audio DiT, not a standard image-model architecture — preprocess_conv [2304, 2304, 1], postprocess_conv [128, 128, 1], latent_conditioners [2048, 4096, 3] — with a separate timestep-feature path and a small cond_layer_logits gate.
Kept at high precision rather than quantized:
preprocess_conv, postprocess_conv, latent_conditioners,
timestep_features, to_timestep_embed, cond_layer, proj_out, rope
The conv layers and the conditioning path are where a music model has the least tolerance for error — the timestep embedding steers every denoising step, and the latent conditioners carry the structure the audio is built on. llama-quantize does not preserve these automatically; they are restored afterward and verified per file.
Every file was checked after conversion to confirm those layers are stored at F16/F32 and that the tensor count and architecture metadata are intact.
Credits
Quantized by RealRebelAI · GitHub · X