In this study, I explore the application of the Mamba architecture, a variant of Structured State Space Models (SSMs), to image segmentation and classification tasks. I benchmark the models against state-of-the-art architectures such as ResNet, U-Net, SegFormer, and UltraLight VM-UNet. My work introduces several novel Mamba-based models, including UMambaBot-PP and variations of UL-VM-UNet, designed to balance performance with computational efficiency. Using the Brain MRI segmentation dataset, I demonstrate that my models achieve competitive Dice scores while maintaining low parameter counts. Additionally, I apply the Vision Mamba (ViM) model to the CIFAR-10 classification task, showcasing its ability to deliver comparable accuracy to the Vision Transformer (ViT) with reduced training and validation times. My findings highlight the potential of Mamba-based models as efficient and effective alternatives for vision tasks, offering insights into their capabilities and advantages. Further research will focus on optimizing these architectures for broader applications and enhanced parameter efficiency.