It runs surprisingly not too slow with a Q2_K quant. Quality is pretty bad though with that quant. If someone has enough memory and disk space to make an iq3_xxs quant of this model, that would probably be a lot better quality than the Q2 while fitting in 16 GB of VRAM.

ggml_vulkan: 0 = AMD RADV RENOIR (ACO) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none

sd --diffusion-model flux2-dev-Q4_K_M.gguf --vae flux2-vae.safetensors --llm mistral-small-3.2-24b-instruct-2506-q4_k_m.gguf --cfg-scale 1.0 --sampling-method euler --steps 8 --offload-to-cpu --vae-on-cpu --diffusion-fa -W 512 -H 1024 -p 'a red apple on a wooden table. Text across the bottom of image reads : sdcpp Flux2 DEV ' --seed 10