When Parts are Greater Than Sums: Individual LLM Components Can Outperform Full Models