Eye diseases pose a significant global health challenge, particularly affecting developing regions like Bangladesh where access to specialized care is limited. This paper presents an advanced multi-model approach to eye disease detection, utilizing a comprehensive dataset of 16,242 expertly labeled fundus images from Bangladeshi healthcare institutions, spanning ten distinct ocular conditions. We implement and evaluate three categories of classification methods: traditional machine learning algorithms (SVM, LR, DT, RF), pre-trained deep learning models (VGG19, ResNet50, MobileNetV2, EfficientNetB4), and vision transformers (ViT-B/16, DINO, Swin). Our experimental results demonstrate the superiority of the Swin Transformer, achieving 0.9100 accuracy score across all metrics, substantially outperforming conventional approaches. The comprehensive analysis includes training curve evaluation showing steady improvement from 70.99 to 93.83% training accuracy over 22 epochs, and detailed confusion matrix analysis revealing exceptional performance for Retinal Detachment and Diabetic Retinopathy, while identifying classification challenges for Myopia (84.4%) and Glaucoma (87.4%). This study advances the development of automated diagnostic tools that can enhance early disease detection and improve healthcare delivery in developing regions through robust multi-class eye disease classification.