Load Molmo Model

Load Molmo Model is the pack's third dedicated loader, and it loads the one model that does something the other two can't: Molmo points at things and counts them, outputting structured coordinates instead of just words. It filters ComfyUI/models/LLM down to folders whose config.json declares the MolmoForCausalLM architecture and returns a VISION_MODEL for the Molmo generation node.

One input (model_name, dropdown), one output (VISION_MODEL). That's the surface - but there's a subtlety under it worth knowing before you run it.

The trust_remote_code thing

Unlike the other loaders, this one loads its processor with trust_remote_code=True hardcoded in. Molmo ships custom modeling code that standard transformers won't load without it, so the author turns it on for you. That's a real tradeoff: running custom code from a model repo is exactly the class of risk that has burned ComfyUI users before (there's a reason people are skittish about trust_remote_code in this ecosystem). The author's own take: "I think Molmo looks safe, though." It's a judgment call - Molmo is a legitimate open-weights model from the Allen Institute - but you should know the toggle isn't even an option here; it's always on.

Model placement

Molmo isn't in the pack's own quantized builds, so you source it yourself. Model folder goes in ComfyUI/models/LLM with the standard layout:

ComfyUI/models/LLM/molmo-7B-D/
├── config.json
├── model.safetensors
└── ...

Molmo is a 7B model, which is friendlier to normal GPUs than the 11–12B siblings. Check config.json shows MolmoForCausalLM as the architecture or this loader will never list it.

Install

ComfyUI Manager search: ComfyUI-PixtralLlamaVision, or:

cd ComfyUI/custom_nodes
git clone https://github.com/SeanScripts/ComfyUI-PixtralLlamaMolmoVision

Requirements: transformers >= 4.45.0, accelerate, bitsandbytes, torchvision. Windows portable:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-PixtralLlamaMolmoVision\requirements.txt

Restart ComfyUI after installing.

Troubleshooting

• Empty dropdown - model not in models/LLM or architecture isn't MolmoForCausalLM.
• Custom-code errors at generation - Molmo needs its own code; if you see loader failures, update transformers. This pack needs 4.45+.
• Not what you expected for OCR or comparison - the README says Molmo is surprisingly bad at bounding boxes; it's built for pointing, counting, and focused questions ("what color is her hair?"). Use the right model for the job.

Load it, feed it to Generate Text with Molmo, and the whole pack's most distinctive feature - coordinates instead of descriptions - becomes available.