RunningHub MiniMax H3 Model Loader

MiniMax H3 is the 33B open-weights model that finally put native stereo audio in an open video generator - it treats text, image, video and audio as one input context and spits out 4–15 second clips at up to 2K/24fps with sound baked in, not bolted on. This node is the front door to that model inside ComfyUI. If you're building an H3 workflow in 2026, this is the loader to use; everything else in the pack is the legacy path.

The name is honest about the payoff: this is a "RunningHub" pack, but it doesn't call the RunningHub API, needs no key, and doesn't spin up an SGLang server. The whole model runs in your ComfyUI process on your own weights. The "RH" is just the author's brand - RunningHub is a cloud GPU host, and this pack is their local-first take on the model.

What it loads

The pack splits H3 into three components you load separately - DiT, text encoder, and a dual VAE - and this node handles the DiT, the big transformer that does the actual denoising. The bundle of INT8 ConvRot weights lives at ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (roughly 95 GiB, so budget 110+ GiB of disk before you start). INT8 ConvRot is the pack's own quantization scheme: int8 weights with a Hadamard rotation trick to keep quality, which is what makes a 33B omni model fit on a realistic card at all.

The three inputs that matter:

• transformer_path - pick which DiT you want: MiniMax-H3-FL2VA-int8_convrot.safetensors for the first-frame / first+last-frame (keyframe) model, or MiniMax-H3-Ref2VA-int8_convrot.safetensors for the multimodal-reference model. The node reads the partition off the filename, so there's no separate dropdown to get wrong.
• adapter - optional turbo LoRA, default none. The H3 weights are trained for a lot of steps; the turbo LoRAs cut that to 4 or 8, which is the difference between waiting forever and actually getting a clip done.
• lora_strength - 0 to 4, default 1.0. The tooltip's advice is the right advice: leave turbo at 1.0 and don't fiddle until you know what you're doing.

The single output is h3_model, a MINIMAX_H3_DIRECT_MODEL handle that wires into the sampler (and the experimental Frame Rate node if you go down that path).

How it works

Under the hood this is a _MergedPartitionLoaderMixin over the direct loader - it resolves the model root, picks the right partition from the filename you chose, and loads the DiT in-process with the attention backend following ComfyUI's current setting. The key difference from the older granular loaders in this pack: you don't pick a model_root or a dtype, and you can attach a LoRA at load time. It also tries both partitions as a fallback if a filename doesn't obviously say which one it is, which is a nice safety net.

Install

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 ComfyUI. ComfyUI Manager works too - search the pack title and install. You'll also want ComfyUI 0.27+ (0.28+ recommended), a CUDA PyTorch build, and the model bundle itself:

cd /path/to/ComfyUI
pip install -U huggingface_hub
hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT

China users can grab the same bundle from ModelScope instead.

The gotchas

Two things bite everyone. First, the license: MiniMax's H3 Community License excludes the US, EU, UK and South Korea from the "Applicable Territory" - running the local weights there is unlicensed, and the pack's own code is Apache-2.0 but the weights are not. Second, the download is enormous and the model is heavy; don't start this on a 12GB card and expect a party. The INT8 ConvRot quantization is exactly the community-weights path people were waiting for at launch, but it's still a 33B omni-modal model.

One more: if an old workflow throws "Node type not found" for MiniMaxH3ModelLoader, that's the pre-RH naming - run python3 tools/migrate_workflow.py old_workflow.json --in-place and it'll be re-pointed at this node.