One of the major complications of diabetes, Diabetic Retinopathy, significantly contributes to visual impairment and blindness. Early and accurate classification of DR stages is vital for guiding prompt intervention and minimizing the risk of advanced retinal damage. However, existing deep learning models face challenges such as limited feature representation, high computational complexity, and class imbalance in DR datasets. To address these issues, this paper presents HIM-Net a DL architecture for multiclass DR classification using fundus images from the APTOS 2019 dataset. HIM-Net combines Inception modules for capturing multi-scale spatial features with lightweight MobileViT blocks that integrate convolutional and Transformer-based representations. Inverted residual blocks are incorporated to enhance feature reuse and reduce model parameters. To mitigate the effects of severe class imbalance, SMOTE is applied at the image level. Additionally, a pre-processing pipeline including resizing and normalization are implemented to standardize input data. Experimental results demonstrate that HIM-Net surpasses several recent models, achieving an accuracy of 95.33% on the APTOS 2019 dataset.