Installing FLUX.2 Klein with ComfyUI: Verified Checklist

By Dominic Marrocco, 42 UK Research. Technical claims below are bounded to the linked primary sources.

Australian installation edition · evidence review: 5 August 2026

Start with the model card, not the old 2 GB claim

The earlier version of this page described FLUX.2 klein as a 2 GB model. That was not supported by the publisher’s documentation. Black Forest Labs says the 4B base model fits in about 13 GB of VRAM and the 9B release fits in about 29 GB. The releases also have different architectures, workflows and licences. Treat their current cards as the source of truth before downloading weights.

Installation checklist

1. Select the model and licence. The cited 4B base weights are Apache 2.0; the cited 9B weights use the FLUX non-commercial licence. Do not assume the 9B terms match the 4B terms.
2. Select a supported ComfyUI path. The official ComfyUI repository links the maintained desktop, portable and manual installation options for supported platforms.
3. Check current requirements. Use the official requirements page for the current Python, PyTorch, operating-system and accelerator guidance. Do not copy an old version number from a blog post.
4. Use the model-specific instructions. Follow the relevant Black Forest Labs card for its required files and example workflow. A generic checkpoint-loader recipe is not a substitute for model-specific documentation.
5. Record the installed revisions. Save the ComfyUI commit or release, model revision, workflow JSON and launch flags so a result can be reproduced.

Memory errors after installation

Use ComfyUI’s official troubleshooting sequence rather than guessing. Its documented options include --lowvram, smaller image dimensions or batches, --preview-method none, and reserved-VRAM controls. Change one variable at a time and record it. An application launch is not proof that a full workflow completes at a claimed memory level.

Where the reusable evidence lives

Use the installation matrix to compare supported setup paths, the workflow library for workflow trust levels, and the low-VRAM benchmark hub for the measurements required before publishing a hardware claim.