Let’s talk about FLUX.1 Kontext dev, this open-source image editing model with 12B parameters. It’s not text-to-image—you give it an existing image plus a command to modify it, like “change the jacket to red” or “swap the background to a forest”—that kind of local edit. VRAM-wise, I tested it: the FP8 checkpoint runs on about 12G, and it’s nearly twice as fast as the original BF16.
If you push it further with NF4 quantization, you can squeeze it down to 7G, but the image quality takes a noticeable hit—don’t bother unless your card is really tiny. The T5 text encoder also eats an extra 6 to 9G of memory, so don’t forget to factor that in. The official API costs four cents per image, so you’d need to edit tens of thousands locally just to break even.
Bottom line: if you’ve got 12G+ VRAM, go local with FP8—set it up in twenty minutes. If your card can’t handle it and you’re not batch-editing images every day, just stick with the API and save yourself the hassle.