We run a 4B model for agent tool calls on laptops. What bit us most was how it behaves around tools. A small model takes the first empty tool search as the truth and gives up, a big one just searches again. Telling it in the prompt didn't help, so we check for that in code now.