black-forest-labs/flux-3-image-klein is the reference inference code repository for FLUX 3 Image Klein, released by Black Forest Labs on October 15, 2026. It is the official implementation for running the 5.6B open-weight image model locally, written in Python and licensed under Apache-2.0. The formal release version is not listed in FACTS.
The model supports generating images from text, editing using reference images, and instructing screen layout with bounding boxes. Editing accepts up to 10 reference images. The repository covers both a distilled version that runs 16 steps without guidance and a 50-step model based on CFG 3.5. According to the README, the distilled version is roughly 6 times faster than the full model, and 1MP image generation times on an H200 are about 3 seconds and 12 seconds respectively. The code includes FP8 and Blackwell-specific NVFP4 quantization paths, and also documents an integer 8-bit distribution for 8GB-class graphics cards along with how to run it in ComfyUI.
However, the roughly 6x speed gap and the image generation times are measurements taken by Black Forest Labs in its own environment. The README's Elo-latency Pareto graph likewise rests on internal evaluation and is hard to reproduce externally. The performance figures are therefore better regarded as measurements reported by the repository rather than independently verified comparison results.