Runtime optimizations alone provide a 2–7× speedup with bitwise-exact outputs. SFew-step distillation provides another 4–5× speedup