You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This repository contains a 8-bit quantization of black-forest-labs/flux-3-image-klein by Black Forest Labs, distributed under the FLUX Kommunity License. The license allows non-commercial use, and commercial use of outputs only for Qualifying Users (gross annualized revenue under US$5M) who meet its conditions. Read LICENSE.md before you continue.

FLUX 3 Image Klein: MLX 8-bit

FLUX 3 Image Klein (distilled, 16 steps, no CFG), quantized to 8 bits for Apple Silicon. It runs with mflux: text-to-image, and editing with 1 to 10 reference images.


Source	black-forest-labs/flux-3-image-klein at revision `8454017f`
Quantization	8-bit transformer, text encoder and VAE (`mflux-save -q 8`)
Size on disk	9.2 GB (the BF16 download is about 20 GB)
Peak memory	22.8 GB at 1024² (M4 Pro)
Defaults	16 steps by default, no guidance

Other variants: cguldogan/flux3-klein-mlx-4bit, cguldogan/flux3-klein-base-mlx-8bit.

License and notice

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

This is a modified version of the FLUX Model. Can Guldogan quantized the weights of black-forest-labs/flux-3-image-klein (revision 8454017f9ebaada30c998fccd9f67a80c7580f68) to 8-bit for Apple MLX with mflux-save. It is not an official product of Black Forest Labs, and Black Forest Labs has not endorsed, approved or validated it.

The weights stay under the FLUX Kommunity License. Your rights to use them come directly from Black Forest Labs under that license. The Qwen3-VL-4B text encoder weights are under the Apache License 2.0 (text_encoder/LICENSE). Section 2.e of the license requires Qualifying Users who use outputs commercially to apply content filters. This package contains no safety classifier.

Use with mflux

FLUX 3 support was merged into mflux in PR #878. Until it is in a release, install mflux from main:

uv tool install --force git+https://github.com/mflux-community/mflux
hf auth login   # this repository is gated: accept the license on this page first

Text-to-image:

mflux-generate-flux3 \
  --model cguldogan/flux3-klein-mlx-8bit \
  --prompt "A small wooden sailboat on a calm alpine lake at dawn, mist over the water, snowy peaks behind." \
  --width 1024 --height 1024 --seed 42 --output boat.png

Edit with one or more reference images:

mflux-generate-flux3-edit \
  --model cguldogan/flux3-klein-mlx-8bit \
  --image-paths boat.png \
  --prompt "The same wooden sailboat on the same alpine lake, now at night under a full moon, stars above the snowy peaks." \
  --seed 42 --output boat_night.png

Add --low-ram to free the text encoder before the denoising loop. Keep flux3-klein in the name of any local copy of this repository: mflux reads the model type from the name.

Prompting

Write full scene captions: framing, subject, background and lighting. For an edit, describe the whole target image with the change folded in, not only the instruction. Black Forest Labs' prompt upsampler is not part of mflux. For layout control, append bounding boxes to the prompt as described in the FLUX 3 model card.

Validation

The mflux port was checked against Black Forest Labs' reference code with exported noise and text embeddings. In float32, the decoded image matches at 94.7 dB PSNR, and at 58.9 dB end to end. The quantized images look the same as BF16 to the eye (8-bit: 44.5 dB PSNR against BF16, 4-bit: 25.2 dB with small changes in detail).

Downloads last month: -

MLX

Hardware compatibility

Quantized

Model tree for cguldogan/flux3-klein-mlx-8bit

Base model

black-forest-labs/flux-3-image-klein

Quantized

(13)

this model