Diabetic retinopathy (DR), is a complication resulting from the disease that can lead to blindness if not detected early. Recently, many classification systems for diabetic retinopathy have been developed. However, several problems were found, namely, the classification results in certain classes still have less than optimal accuracy values, the lack of in-depth analysis for the results, and the overall accuracy that can still be improved. In this work, we experiment by evaluating and combining new deep learning models such as EfficientNet, EfficientNetV2, LCNet, MobileNetV3, TinyNet, and FBNetV3 using ensemble stacking techniques with four different meta-learners: decision trees, logistic regression, ANN, and SVM to provide better accuracy in classifying the severity of diabetic retinopathy. Our work offers satisfactory classification results on the APTOS 2019 dataset with training, validation, testing, and F1 score accuracy of 96.56%, 95.33%, 84.17%, and 70.16%, respectively.