RunningHub MiniMax H3 FL2VA Qwen3-VL Loader (Direct) (Legacy)

MiniMax H3's text encoder is a Qwen3-VL-32B model - a full-size vision-language model doing the job a CLIP usually does, because H3 needs to fuse text, images, video and audio into one shared conditioning context. This legacy node loads that encoder pinned to the FL2VA partition of the weights. The "FL2VA-fixed" part is honestly a formality: both partitions ship the identical text encoder (the source comments note they're byte-for-byte the same size), so this node and its Ref2VA twin differ only in which weights directory they prefer to resolve.

That's the key thing to know here: there is no functional difference between this and RHMiniMaxH3DirectTextEncoderLoader beyond the partition pin. This is a legacy compatibility node. New workflows should use RHMiniMaxH3TextEncoderLoader - the modern loader that dropped the partition knob entirely because the choice could never matter. If you've got an old FL2VA graph with this node in it, it keeps working, and that's the point.

Inputs

• model_root - weights root, models/MiniMax-H3-INT8-CONVROT (legacy models/MiniMax-H3 accepted).
• dtype - auto, bfloat16, float16, or float32; auto/bf16 recommended.
• text_encoder_path - the encoder file, default qwen3-vl-32b-int8_convrot.safetensors.

Output: one h3_text_encoder handle for the FL2VA encode node.

Why the encoder matters more than this node does

Don't be fooled by how boring a "legacy loader" sounds. The encoder is where H3's omni-modal magic lives from your side of the graph - it's what lets the same prompt carry an image reference and an audio reference and a sentence of direction into one conditioning tensor. And it's a 32B model, so it's a real VRAM resident and a real reason the INT8 ConvRot bundle exists. The node itself is trivial: pick a root, pick a dtype, get a handle. The weight is the story.

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, pull the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT, ModelScope in China), and remember the transformers floor in requirements.txt (≥4.57) isn't optional - Qwen3-VL won't load on an old one. License reminder as ever: MiniMax H3's Community License excludes the US, EU, UK and South Korea from running the local weights.