Black Forest Labs, the Freiburg image-model company, released FLUX.3 Image Klein on 15 October in two checkpoints. The distilled version defaults to 16 steps with no classifier-free guidance and is built for speed, while the base version defaults to 50 steps and is the one the lab recommends for fine-tuning. The model does text-to-image, multi-reference image-to-image and bounding-box layout control, and single-file weights are published in bf16, FP8 and NVFP4 for smaller cards.
The size is genuinely unclear. The lab’s own announcement gives 5.6bn parameters. The Hugging Face model cards say 6bn. Neither explains the gap, so the figure is best treated as approximate until one of them corrects the other.
The performance claim needs the same caution. Black Forest Labs says the model leads its blind evaluations for text-to-image, multi-reference generation and editing, but those are the lab’s internal tests rather than a published benchmark anyone else can run. Neither the announcement nor the Hugging Face integration post states licence terms, which again leaves the commercial position of these weights unsettled. The announcement is on Black Forest Labs’ site.