5 gigabytes vram in its minimum configuration, but various things can be done to increase that. Quantization and distillation might theoretically reduce resource needs, but that's still small enough to get halfway decent CPU generation time.
It's hard to infer relative performance based on parameter count alone. SD3 and SDXL are quite different architecturally. The only way to really tell is to compare it with examples. Even this lobotomized 2B model seems to perform better on prompt adherence and text compared the base SDXL model, so I think it has potential once fine tuned.
