We make Chain-of-Thought reasoning in large language models (1) more efficient by creating implicit reasoning with lightweight language models; (2) still effective as the implicit reasoning maintains semantic alignment with ground-truth reasoning.