Grading bone marrow cells is essential in hematological diagnostics for differentiating conditions such as leukemia, lymphoma, and myelodysplastic syndrome. Traditional evaluation relies heavily on manual visual inspection, which is subjective and labor-intensive. This paper proposes a hybrid deep learning framework that integrates Vision Transformer and MobileNet to classify five key bone marrow cell types: Hairy Cell, Faggot Cell, Basophil, Karyocyte Stem Cell, and Immature Lymphocyte. Multiple CNN architectures, including ResNet50, VGG19, DenseNet121, InceptionV3, and MobileNet, were benchmarked for comparison. The ViT+MobileNet model leverages ViT’s global self-attention and MobileNet’s efficient local feature extraction, while a Dual-Attention DenseNet variant was introduced to improve feature localization. Hyperparameters were optimized using Bayesian optimization combined with grid search. The hybrid ViT+MobileNet and tuned MobileNet models achieved the highest F1-score of 0.98, outperforming DenseNet121 (0.86), MobileNet (0.91), VGG19 (0.77), InceptionV3 (0.75), and ResNet50 (0.58). The proposed system addresses key challenges in hematological imaging-including class imbalance and morphological similarity-through advanced augmentation and fine-tuning, offering a robust and clinically scalable solution for automated bone marrow cell classification.