Diabetic retinopathy (DR) is a major cause of vision loss and requires accurate and timely diagnosis. This study proposes DenseViT-DR, a hybrid deep learning framework that combines DenseNet121 and Vision Transformer (ViT) for multi-class diabetic retinopathy severity classification using retinal fundus images. The model was evaluated on a Gaussian-filtered retinal dataset derived from the APTOS 2019 benchmark and compared with ResNet-18, VGG16, and a conventional CNN model. Experimental results show that DenseViT-DR achieved the highest performance, with 96.33% accuracy and a Micro-F1 score of 0.963, outperforming ResNet-18 (94.10%), VGG16 (93.45%), and the basic CNN model (90.62%). Furthermore, external validation on an independent clinical dataset collected from Bangladesh Eye Hospital achieved 90.40% accuracy and a Macro F1-score of 0.898, demonstrating strong generalization capability. These findings indicate that DenseViT-DR is an effective and robust framework for automated diabetic retinopathy severity classification.