FL MiniMax Music 3 Loader

MiniMax Music 3 is the closest the open-weights world has come to Suno - a diffusion model that generates real music from a caption and optional lyrics, not MIDI or spectrogram soup. This is the node that starts every Music 3 workflow, and it does the part everybody dreads: the 13.34 GiB first-run download, spread across three verified files, before you hear a single bar.

Here's the thing to understand about that download: it's not lazy. The pack pins a specific revision of the official Comfy-Org/MiniMax-Music-3 repo and hard-codes the sha256 of every file. Each artifact is downloaded with a .part resume, so interrupting a queue mid-download doesn't cost you the progress - you just re-queue and it picks up where it left off, with per-file progress right on the node. That "verified" bit matters more than it sounds: a truncated or corrupted file that quietly fails on load is the classic way a 13 GB download eats your evening.

What you actually get

The node has zero required inputs. It's a pure loader that hands you three outputs:

• model - the diffusion model (minimax_music3_dit_fp16.safetensors, ~4.9 GB), wired into the sampler.
• clip - the MiniMax text encoder (a pruned int8 variant, ~9.2 GB), wired into the CLIP/text-encode side of the workflow.
• vae - the DAV (deep audio VAE, ~216 MB) for turning latents back into audio.

If you've used any ComfyUI workflow, the wiring is familiar: model and clip go toward the sampler, VAE goes to the encode/decode end. It slots into the standard MiniMax Music 3 inference workflow that ComfyUI ships support for. One honest gotcha: this pack's loader needs a current ComfyUI. If your install predates MiniMax Music 3 support, it throws a specific error - "This ComfyUI version does not support MiniMax Music 3" - rather than silently loading garbage. That's a feature. Update ComfyUI and move on.

The two knobs you might touch

Both optional, both tucked in the advanced section, and both exist to squeeze onto smaller VRAM:

• weight_dtype - default follows ComfyUI's model policy; fp8_e4m3fn, fp8_e4m3fn_fast, and fp8_e5m2 load the diffusion model in fp8. If you're on a 12–16 GB card and OOMing, fp8 is the lever. _fast also turns on fp8 optimizations.
• clip_device - default or cpu. Shoving the 9 GB text encoder onto the CPU trades some speed to free the GPU.

Install

ComfyUI Manager: search FL MiniMax Music 3. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxMusic3.git
cd ComfyUI-FL-MiniMaxMusic3
pip install -r requirements.txt

Then restart ComfyUI. The pack's own Python requirement is trivial (requests - the real dependencies ride along in ComfyUI itself), but note the README asks for Python 3.10+ and a current ComfyUI build. Model files download into the normal ComfyUI model folders (diffusion_models, text_encoders, vae) and only download when you actually queue the node - so don't panic when the first queue sits there for a while.

Where people get burned

• Not updating ComfyUI first. The loader error is the tell; it won't work around an old core.
• Ignoring disk space. 13.34 GiB for the base set, plus more if you add the Audio VAE loader or MOSS preprocessing. Check free space before your first queue.
• Reading the 13.34 GiB as "the whole pack." It's the base model set. The training side later asks for another ~20 GiB of backend plus model storage. This loader is the cheap part.