That's mostly correct, but I believe the INT8 support on M5 changes the equation pretty drastically. I also believe that diffusion models generally output _better_ results if they have shaping for code (they often will miss delimiter matching inside of a canvas).