CED

A neural network architecture design proposed by DeepSeek that splits model layers into an encoding (reading) phase and a decoding (generation) phase, assigning differentiated parameter activation volumes to each phase. It further optimizes inference efficiency on top of the MoE framework.