Early diagnosis is crucial for the effective treating eye diseases and the preventing loss of vision. This study presents a novel comparison between two deep learning models, Inception V3 and SqueezeNet, for the classification of eye illnesses, specifically glaucoma, uveitis, diabetic retinopathy, and cataracts. While existing research has explored deep learning for eye disease detection, few studies provide a direct, in-depth performance evaluation between a high accuracy but computationally intensive model like Inception V3 and an efficiency-driven model like SqueezeNet. Both models have been trained and evaluated on a large dataset of retinal images in order to assess how well they performed in terms of accuracy, sensitivity, precision, recall, and F1-score, among other metrics. The results demonstrate that Inception V3, in comparison with SqueezeNet, delivers greater overall accuracy and precision despite its complexity. This study demonstrated SqueezeNet and Inception V3 learn from training data with 91.48 and 91.07% accuracy. SqueezeNet enhanced generalization by raising validation, training, and test accuracy to 88.01%, 95.60%, and 85.43%, respectively, after training. Inception V3 has 99.23% training accuracy, 92.79% validation, and 90.87% test accuracy. This work uniquely contributes to the field by: (i) providing a direct performance trade-off analysis between a computationally intensive and a lightweight deep learning model for ophthalmic diagnostics; (ii) demonstrating the impact of fine-tuning on improving generalization across both models; (iii) addressing clinical applicability by offering insights into model selection based on diagnostic priorities—Inception V3 for high-accuracy scenarios and SqueezeNet for real-time, resource-efficient applications. This research clarifies model complexity-performance trade-offs to improve clinical AI integration decision-making and promote the deployment of new diagnostic technologies to improve early diagnosis of diseases related to the eye.