RunningHub MiniMax H3 Model Loader (Direct) (Legacy)

This is the "direct" DiT loader for MiniMax H3, and yes, it says "(Legacy)" in the title - but "legacy" here doesn't mean "broken," it means "superseded." The pack keeps every old node registered so workflows that already exist keep loading and running; it just hides them from search and the node tree so new users don't trip over a dozen loaders. If you're starting from scratch, reach for RHMiniMaxH3ModelLoader instead. If you opened an existing H3 workflow and this node is in it, it'll work exactly as it did before.

What makes it "direct" is the thing that makes the whole pack worth knowing: the DiT loads straight into your ComfyUI process. No SGLang server, no Diffusers pipeline, no API key. MiniMax H3 is a 33B omni-modal video model with native stereo audio, and "direct" means you're running those weights yourself, on your own hardware.

What it loads and what you set

This loader takes an explicit model root and component path - that's the "direct" contract. Instead of the newer loader guessing your partition from a filename, you name everything:

• model_root - the weights root, models/MiniMax-H3-INT8-CONVROT by default (the old models/MiniMax-H3 layout still works too). All three components - DiT, text encoder, VAEs - must come from the same root.
• dtype - auto, bfloat16, or float16. auto/bfloat16 is the recommendation; the runtime keeps the fp32 layers H3 specifies even in bf16 mode.
• transformer_path - the explicit DiT file, default MiniMax-H3-FL2VA-int8_convrot.safetensors. Unlike the newer loader, it will not auto-swap between quantized/BF16 weights - you say what you want.

The optional attention_backend is where this node still has an edge: auto (follow ComfyUI), sdpa (PyTorch), sage (SageAttention if installed), or ck (Comfy Kitchen INT8 - the same kernel as --use-ck-attention, and it hard-errors if that's not available). If you've got SageAttention working, forcing sage is a real speedup on a model this size.

Output is a single h3_model handle that feeds the Dual Sigma Sampler.

Install

Same pack install as everything else here:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, then make sure the ~95 GiB INT8 ConvRot bundle is in ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (download with hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT, or the ModelScope mirror in China).

When you'd actually use this one

Honestly: rarely, and that's fine. It's a fine node - explicit roots, sensible defaults, the ck attention option is genuinely handy if you run comfy-kitchen. But the recommended loader adds LoRA support and partition auto-detection while removing two knobs you never needed to touch. Use this if you're maintaining an old workflow or you specifically want to pin a dtype and attention backend by hand; use the new loader for anything greenfield. Don't be alarmed when you don't find it in the node menu - search shows the modern nodes, and old graphs keep resolving to these. That's the migration story the pack intends, and it works.

The license reminder applies to everything in this pack: MiniMax H3's community license excludes the US, EU, UK and South Korea from running the local weights, and those terms cover the model, not the node code.