FLUX.2 [klein] is a smaller branch of the FLUX.2 lineup designed to make interactive visual intelligence practical on everyday hardware, rather than just big servers. The family centers on two sizes, 4B and 9B parameters, each offered as a fast distilled version and a more flexible base model. Both can handle text to image, image to image editing, and multi reference generation in a single unified model, which helps keep workflows simple while still supporting complex compositions and style transfer. The 4B variant is fully open under Apache 2.0 and fits in roughly 13 GB of VRAM, so it targets RTX 3090 or 4070 class cards and is tuned for sub second latency.
The 9B model aims for higher quality and still keeps latency low enough for interactive use, with distilled checkpoints reaching around half a second to a couple of seconds per image on high end GPUs. Quantized FP8 and NVFP4 versions, developed with NVIDIA, further reduce VRAM and speed up inference, which broadens the range of machines that can run the models comfortably. Overall, FLUX.2 [klein] is positioned as a practical bridge between heavyweight frontier image models and lightweight tools that sometimes compromise too much on image fidelity
