MiniMax H3 NF4 Text Encoder Loader / NF4 文本编码器加载

MiniMax H3's text side is a Qwen3-VL encoder, and in the pre-quantized DiffSynth NF4 form it's a 15.32 GB file that ComfyUI can't just load with the stock CLIPLoader. Same story as the pack's model loader, one level down: the DiffSynth checkpoint stores Linear weights as flattened BitsAndBytes tensors with their shapes hidden in .quant_state.bitsandbytes__nf4 metadata, and it uses DiffSynth's own weight prefixes (model.language_model.*, model.visual.*) that don't match what ComfyUI expects. This loader restores the NF4 quant state, converts the prefixes into ComfyUI's layout, and hands you a normal CLIP output.

What you need

Drop this into ComfyUI/models/text_encoders/:

| File | Size | |---|---| | minimax-h3-text-encoder-nf4.safetensors | 15.32 GB |

Select it in the encoder_name combo, connect the CLIP output where the standard H3 text encoder loader was connected, and the rest of the workflow stays untouched. It's one of the four NF4 loaders in this pack that together replace the stock UNET/CLIP/VAE loaders with the DiffSynth NF4 stack.

BitsAndBytes is required and it's CUDA-only:

python -m pip install "bitsandbytes>=0.50.0"

If it's missing, the loader stops with a message telling you exactly that - it doesn't half-load.

Why use the NF4 encoder at all

Because the encoder is a big chunk of H3's footprint. The full Qwen3-VL encoder is the kind of weight that squeezes whatever's left of your VRAM, and NF4 shrinks the download and the resident size. Same honest caveat as the other NF4 loaders, though: it dequantizes for computation, so smaller doesn't automatically mean faster. The pack's validation covered the Qwen3-VL NF4 encoder in isolation, so this path is tested - but if you have room for the Comfy-Org int8_convrot encoder, that's the faster route on the machine they measured.

Install

Part of the MiniMax-H3 Tools pack (ComfyUI 0.30.0+). Search "MiniMax H3 Tools" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git

Restart ComfyUI and find MiniMax H3 NF4 Text Encoder Loader under MiniMax H3. It fails explicitly on FP4 files, missing quant sidecars, non-H3 models, and incomplete downloads (.incomplete / .aria2log artifacts are not model files). One more thing worth knowing before you download anything: H3 weights are licensed under the MiniMax H3 Community License, which excludes the US, EU, UK and Korea from running the local weights - that applies to the encoder file too.