DeepSeek-V2-Lite

Researchers propose DraftExpert, an expansion-aware self-speculative decoding framework for expert-offloaded Mixture-of-Experts (MoE) inference on end devices, achieving 1.45x average decode throughpu…