Proposed architecture for inferencing sparse MOE models increasing Active parameters using layered + linear decay. Succinct reasoning without any model training or fine tune. [p]