2026-10-15

Black Forest Labs Releases FLUX 3 Image Klein, a 5.6B-Parameter Open-Weights Image Model With Native 4K, Bounding Boxes and Pixel-Perfect Editing That Runs on One Consumer GPU

AIModelsOpen Source🌍 Europe

Black Forest Labs has released FLUX 3 Image Klein, an open-weights image model that the lab says has "the quality of a large frontier model" while staying customizable and fast. At 5.6B parameters it runs on a single consumer GPU, and it supports native 4K output, bounding boxes and pixel-perfect editing, per BFL's announcement.

FLUX 3 Image Klein open-weights announcement collage from Black Forest Labs, with labels for image, true to life, text, style, detail and bounding box

From Black Forest Labs' announcement.

What it adds

The three headline features line up with what FLUX 3 Image introduced on October 1, and Klein brings them to an open-weights model small enough for local hardware, according to BFL's announcement.

  • Pixel-perfect editing. FLUX 3 Image was pitched on multi-turn edits that change only what was asked for and leave every other pixel untouched. Klein's announcement carries the same promise.
  • Bounding boxes. FLUX 3 Image lets users place each object and each block of text with a labeled box instead of hoping a prompt lands them in the right spot. Klein's visual shows the same bounding-box grid.
  • Native 4K. FLUX 3 Image generates in up to 4K, and Klein lists native 4K as well.

The announcement collage also labels the model's strengths as image, true to life, text, style and detail.

The size is the other half of the news. A 5.6B model that fits on one consumer GPU can be fine-tuned and run locally by individuals and small teams, which is what "customizable and fast" points at.

Context

When BFL launched FLUX 3 Image, it offered Commercial Weights to companies running generation at scale and promised an Open Weights version in the coming weeks. Klein arrives two weeks later as the open-weights release. The wider FLUX 3 family, covering image, video and robot action, was introduced in the July FLUX 3 post.

It is the second open image model of the month. Alibaba's Qwen team released the 7B Qwen-Image-2.1-Turbo as open weights on October 9, which generates 2K images in 8 denoising steps. Datapoint's image Elo table had the top seven image models all closed in August.

Read next