MiniMax Music 3 Models Loader ⚡

MiniMax Music 3 is a three-piece pipeline: an autoregressive conditioner that plans the song from your caption and lyrics, a flow-matching DiT that turns that plan into audio latents, and a DAV decoder that renders the final waveform. This loader pulls all three from your own models folders and hands them back as plain comfy MODEL / CLIP / VAE - which means everything downstream composes with stock nodes. No bundled model code, no runtime downloads.

The three dropdowns

• transformer_name - the flow DiT, from models/diffusion_models. The tooltip names minimax_music3_dit_int8_convrot.safetensors as the reference file; fp16/bf16 and GGUF quants are supported too.
• text_encoder_name - the pruned AR text/audio encoder, from models/text_encoders. It ships as minimax_music3_text_encoder_pruned_int8_convrot.safetensors and carries its tokenizer_json embedded, so nothing extra to fetch. A prepared GGUF works here as well.
• dav_name - the decoder, minimax_music3_dav.safetensors, from models/vae.

Download the Comfy-Org/MiniMax-Music-3 repo and place its split files into those three folders. That's the whole install - no HuggingFace runtime step.

What the loader does behind the scenes

Two things are worth knowing. First, it composes ComfyUI's native MiniMax Music 3 implementation - no model code is vendored in this pack, so you're on core's architecture and its bug fixes. Second, the quantization story: the native ConvRot INT8 safetensors go through Comfy's optimized mixed-precision path, while GGUF weights stay packed through the pack's ops and dequantize on the fly. That "stays quantized" behavior is the GGUF selling point across this whole pack, and it matters here because the DiT is the heaviest piece - a GGUF quant of it is the difference between fitting and OOM on a 12GB card.

The outputs

• model (MODEL) - the flow DiT, patched and ready to sample.
• clip (CLIP) - the pruned AR conditioner, ready for CLIPTextEncode-style prompting.
• vae (VAE) - the DAV decoder, ready to turn latents into AUDIO.

Wire all three into MiniMax Music 3 Audio Generate and you're done: caption + lyrics → 44.1 kHz stereo music.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/MiniMax Music. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Only gguf is strictly required for inference; the rest of requirements.txt serves the pack's other optional nodes.

Common issues

The dropdowns are populated from your folders, so if one is empty the file isn't where the tooltip says it should be - recheck models/diffusion_models, models/text_encoders, and models/vae (the README's file table is the source of truth for names). Mixing a GGUF DiT with an int8 text encoder is fine; just don't expect the GGUF piece to take Comfy's mixed-precision ConvRot path - it stays quantized through the pack's own ops instead, which is normal and works. If a load fails, the pack ships tools/smoke_minimax_music3.py, which validates all three checkpoint load paths without generating a song - handy for separating "file problem" from "sampling problem."