nForce101/flux2-klein-4b-q4

FLUX.2 [klein] 4B — MLX 4-bit (Pixai mirror)

A pinned mirror of the 4-bit MLX pack of black-forest-labs/FLUX.2-klein-4B, used by the Pixai macOS app for on-device generation and multi-image editing. Weights are byte-identical to mflux-community/flux2-klein-4b-mflux-q4 at revision 794cd159538149ad9830848508c31f0ea7088e58. Not an official Black Forest Labs release.

Folder	Contents	Size
transformer/	FLUX.2 klein 4B transformer, MLX 4-bit (group size 64)	2.18 GB
text_encoder/	Qwen3-4B text encoder, MLX 4-bit	2.26 GB
vae/	FLUX.2 VAE	0.17 GB
tokenizer/	Tokenizer used by the mflux loader	11 MB
bfl-config/	Original model configs and full tokenizer from BFL (revision e7b7dc27)	14 MB

pixai-model.json lists every file with its size and SHA-256; Pixai verifies each download against it.

Use with mflux

mflux-generate-flux2-edit --model <this folder> --base-model flux2-klein-4b --low-ram \
  --steps 4 --guidance 1.0 --image-paths person.png hat.png \
  --prompt "Make the woman from image 1 wear the hat from image 2"

Measured on a 16 GB MacBook Air (M4), 1024², 4 steps, --low-ram: text-to-image ≈85 s, two-reference edit ≈3–4.5 min, peak memory ≈7.7–9.9 GB.

License

Apache License 2.0 — see LICENSE.md and NOTICE. The original model's license applies unchanged.