That's a cool write up! I suspect a lot of this will have to do with quantization, how much quality is getting lost, and how each model behaves under different quants.