I replaced my local LLM with a model half its size and got better results — And it wasn't about the parameters

I switched from a 20B model to a 9B one, and it was better