RehmanKhalid/flux-3-image-klein-Q4_K_M

FLUX 3 Image Klein — Q4_K_M GGUF

Community quantization of the distilled black-forest-labs/flux-3-image-klein transformer (5.6B). The BF16 Comfy checkpoint was converted with molbal/ComfyUI-GGUF to architecture flux3, file type MOSTLY_Q4_K_M.

This is a modified derivative. It is not an official Black Forest Labs product.

File	Quant	Size
flux3_image_klein-Q4_K_M.gguf	Q4_K_M	2.91 GB

270 weight tensors are Q4_K. 87 small tensors stay F32. emb_in.image.weight and emb_in.image_cond.weight stay BF16 because their inner dimension is below the 256-value quant block.

ComfyUI

Use an up-to-date ComfyUI build with FLUX 3 support and the molbal ComfyUI-GGUF loader (UnetLoaderGGUF). This file is only the diffusion transformer.

Component	File	Folder
Diffusion transformer	flux3_image_klein-Q4_K_M.gguf	models/diffusion_models/ or models/unet/
Text encoder	qwen3vl_4b_bf16.safetensors or qwen3vl_4b_int8_convrot.safetensors	models/text_encoders/
VAE	flux2-vae.safetensors	models/vae/

Distilled sampling: Euler / simple, 4–16 steps, CFG 1.0.

Text encoder and VAE come from the official Comfy release, not from this repo.

License

Weights remain under the FLUX Kommunity License. See NOTICE.md. All credit for the model to Black Forest Labs.