Hugging Face's logo

RunPod
/
FLUX.2-klein-4B-mflux-4bit

Text-to-Image
Safetensors
MLX
mflux
flux2
flux2-klein
quantized
4bit
apple-silicon
Model card Files
xet
Community
FLUX.2-klein-4B-mflux-4bit / README.md
NERDDISCO's picture
NERDDISCO
fix: use Runpod branding in display text
0460d11 verified
preview code
|
raw
history blame contribute delete
2.84 kB
metadata 
license: apache-2.0
base_model: black-forest-labs/FLUX.2-klein-4B
tags:
  - text-to-image
  - safetensors
  - mflux
  - mlx
  - flux2
  - flux2-klein
  - quantized
  - 4bit
  - apple-silicon
library_name: mflux
pipeline_tag: text-to-image

FLUX.2 [klein] 4B — mflux 4-bit quantized

4-bit quantized weights of FLUX.2 [klein] 4B by Black Forest Labs, optimized for mflux on Apple Silicon.

Full precision This repo (4-bit)
Size ~8 GB 4.3 GB
Framework diffusers / mflux mflux only
Hardware CUDA / MLX Apple Silicon (MLX)

Quickstart

Install mflux 

pip install mflux

Generate an image 

mflux-generate-flux2 \
  --model RunPod/FLUX.2-klein-4B-mflux-4bit \
  --prompt "A cute robot standing in a field of flowers, digital art" \
  --width 1024 \
  --height 1024 \
  --steps 4 \
  --seed 42 \
  --output output.png

Python usage 

from mflux import Flux2

flux = Flux2(
    model="RunPod/FLUX.2-klein-4B-mflux-4bit",
    base_model="flux2-klein-4b",
)

image = flux.generate_image(
    prompt="A cute robot standing in a field of flowers, digital art",
    width=1024,
    height=1024,
    num_inference_steps=4,
    seed=42,
)

image.save("output.png")

Details

  • Base model: black-forest-labs/FLUX.2-klein-4B (Apache 2.0)
  • Quantization: 4-bit via MLX nn.quantize (group_size=64), created with mflux-save --quantize 4
  • Requirements: mflux v0.16.0+, Apple Silicon Mac
  • Performance: ~11s for 512x512 (4 steps) on M3 Pro 18GB

How this was created 

pip install mflux
mflux-save \
  --path ./FLUX.2-klein-4B-mflux-4bit \
  --model flux2-klein-4b \
  --quantize 4

About FLUX.2 [klein]

FLUX.2 [klein] 4B is a 4 billion parameter rectified flow transformer by Black Forest Labs for fast image generation and editing. It delivers state-of-the-art quality with sub-second inference on consumer hardware.

  • Ultra-fast inference (4 steps)
  • Text-to-image and multi-reference image editing
  • Apache 2.0 — fully open for commercial use
  • Blog post | GitHub

License

Apache 2.0, inherited from the original model.

Credits