Skip to main content
Back

Black Forest Labs' FLUX.2 Klein Runs Sub-Second on a Consumer GPU

Black Forest Labs released a 4B rectified flow transformer that unifies text-to-image and multi-reference editing in under a second on consumer GPUs.

  • FLUX.2 klein 4B is a 4B rectified flow transformer from Black Forest Labs.
  • Sub-second inference, fits in ~13GB VRAM on an RTX 3090 or 4070.
  • Unified text-to-image, image editing, and multi-reference in one checkpoint.
  • Released under Apache 2.0, free for commercial use.
  • Distilled variant plus a Base checkpoint for fine-tuning and LoRA training.
  • Available in Diffusers, ComfyUI, BFL API, and NVIDIA NIM endpoints.

Black Forest Labs released a compact image model that fits on a single consumer GPU and handles both generation and editing within one architecture. FLUX.2 klein 4B is the smallest model in the FLUX.2 family, distilled from a larger base, and licensed under Apache 2.0, which is rare for a state-of-the-art image system.

At 4 billion parameters, klein runs in under a second on an NVIDIA RTX 3090 or 4070 (around 13GB VRAM), using just four inference steps. Its single set of weights covers text-to-image generation and multi-reference image editing, collapsing what used to be separate pipelines into one checkpoint.

Three flavors, one architecture

The release ships in distinct variants:

  • klein 4B (distilled): the default, tuned for sub-second inference with a minimal step count.
  • klein 4B Base: undistilled, preserving the full training signal for fine-tuning, LoRA training, and custom pipelines that need output diversity over raw speed.
  • klein 4B NVFP4: a 4-bit quantized build for NVIDIA hardware that cuts memory further.

Getting it running

The model integrates with Diffusers through a dedicated Flux2KleinPipeline. A minimal call looks like this:

import torch
from diffusers import Flux2KleinPipeline
pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()
image = pipe(
    prompt="A cat holding a sign that says hello world",
    height=1024, width=1024,
    guidance_scale=1.0,
    num_inference_steps=4,
).images[0]