I'm genuinely surprised to see that "thinking" flash-lite is more performant than flash with no "thinking".