To ensure the timeliness of malware variant detection models, traditional machine (deep) learning-based detection methods integrate historical and incremental data and retrain to update detection models. However, this approach often suffers from low training efficiency. Therefore, this paper proposes an incremental learning method based on a neural network smooth aggregation mechanism for detecting malware variants, facilitating the smooth evolution of detection models. The method introduces a training scale factor to prevent the decrement of accuracy in the aggregated incremental model due to small training scales. Experimental results show that the proposed incremental learning method can improve training efficiency while maintaining the accuracy of the detection model compared to the re-training method.