“Homomorphic encryption for language model inference is a potential alternative to local models for privacy, but it is currently in the research stage due to significant performance overhead.”