Fiddler is a fast inference system for LLMs based on Mixture-of-Experts (MoE) architecture at local devices.