FLUX.2-klein-4B โ€” mflux 4-bit (Apple Silicon)

4-bit quantized version of black-forest-labs/FLUX.2-klein-4B, saved with mflux for fast inference on Apple Silicon Macs.

~4 GB on disk (vs 16 GB for the original BF16 weights).
**
9 seconds per image** at 512ร—512, 4 steps on M-series chips.

Sample output

sample

Prompt: "a red apple on a white table", seed 42, 512ร—512, 4 steps

Requirements

pip install mflux

Requires macOS with Apple Silicon (M1/M2/M3/M4). Tested with mflux 0.18.0.

Usage

Python

from mflux.models.flux2.variants.txt2img.flux2_klein import Flux2Klein
from mflux.models.common.config.model_config import ModelConfig

flux = Flux2Klein(
    model_config=ModelConfig.flux2_klein_4b(),
    model_path="ar9av/FLUX.2-klein-4B-mflux-4bit",
)

image = flux.generate_image(
    seed=42,
    prompt="a photo of an astronaut riding a horse on mars",
    num_inference_steps=4,
    width=1024,
    height=1024,
    guidance=1.0,
)
image.image.save("output.png")

CLI

mflux-generate-flux2 \
  --model flux2-klein-4b \
  --model-path ar9av/FLUX.2-klein-4B-mflux-4bit \
  --prompt "a photo of an astronaut riding a horse on mars" \
  --steps 4 \
  --output output.png

With mlx-local-server (OpenAI-compatible API)

# Load the model
curl -X POST http://localhost:8002/v1/models/load \
  -H "Content-Type: application/json" \
  -d '{"model": "flux2-klein", "model_path": "ar9av/FLUX.2-klein-4B-mflux-4bit"}'

# Generate
curl -X POST http://localhost:8002/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{"prompt": "a red apple", "size": "512x512"}'

Model details

Property Value
Base model black-forest-labs/FLUX.2-klein-4B
Quantization 4-bit (mflux format)
Saved with mflux 0.18.0
Architecture Flux2Klein (Qwen3 text encoder, no T5)
Parameters ~4B
Guidance 1.0 (distilled, CFG-free)

Notes

  • This model does not support negative_prompt (distilled model, guidance=1.0).
  • For img2img, pass image_path and image_strength to generate_image.
  • License follows the original FLUX.1-dev non-commercial license.
Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ar9av/FLUX.2-klein-4B-mflux-4bit

Finetuned
(39)
this model