We propose CryptoMoE, the first framework to enable private, accurate, and efficient inference for MoE-based LLMs.