A 9-billion parameter variant of the FLUX.2 image generation model with FP8 key-value cache quantization by Black Forest Labs for memory-efficient inference.