SegFormer3D is a lightweight and efficient Transformer-based architecture designed for 3D medical image segmentation. It addresses the challenges of high computational costs and limited dataset availability in medical imaging by offering a model with 33× fewer parameters and 13× lower GFLOPS compared to current state-of-the-art (SOTA) models. The model uses a hierarchical Transformer to calculate attention across multiscale volumetric features, and an all-MLP decoder to aggregate local and global attention features for accurate segmentation. SegFormer3D preserves the performance characteristics of larger models while being significantly more compact, making it suitable for real-world medical applications. The architecture incorporates overlapped patch merging to reduce information loss during voxel generation and an efficient self-attention mechanism to handle long-range dependencies in 3D sequences. It also uses a mix-ffn module to automatically learn positional cues, eliminating the need for fixed encoding. The decoder uses linear layers for efficient and consistent decoding of volumetric features across diverse datasets. SegFormer3D was benchmarked against SOTA models on three widely used datasets: Synapse, BRaTs, and ACDC. It achieved competitive results, demonstrating high accuracy and efficiency. The model outperformed existing solutions in terms of parameter count and computational complexity, while maintaining strong performance. It is particularly effective in scenarios with limited datasets, where larger models often struggle with generalization and convergence. The paper highlights the importance of lightweight and efficient architectures in medical imaging, emphasizing that they can significantly improve performance without additional pretraining or computational resources. SegFormer3D represents a valuable advancement in 3D medical image segmentation, offering a balance between performance and efficiency that is crucial for practical applications in healthcare.SegFormer3D is a lightweight and efficient Transformer-based architecture designed for 3D medical image segmentation. It addresses the challenges of high computational costs and limited dataset availability in medical imaging by offering a model with 33× fewer parameters and 13× lower GFLOPS compared to current state-of-the-art (SOTA) models. The model uses a hierarchical Transformer to calculate attention across multiscale volumetric features, and an all-MLP decoder to aggregate local and global attention features for accurate segmentation. SegFormer3D preserves the performance characteristics of larger models while being significantly more compact, making it suitable for real-world medical applications. The architecture incorporates overlapped patch merging to reduce information loss during voxel generation and an efficient self-attention mechanism to handle long-range dependencies in 3D sequences. It also uses a mix-ffn module to automatically learn positional cues, eliminating the need for fixed encoding. The decoder uses linear layers for efficient and consistent decoding of volumetric features across diverse datasets. SegFormer3D was benchmarked against SOTA models on three widely used datasets: Synapse, BRaTs, and ACDC. It achieved competitive results, demonstrating high accuracy and efficiency. The model outperformed existing solutions in terms of parameter count and computational complexity, while maintaining strong performance. It is particularly effective in scenarios with limited datasets, where larger models often struggle with generalization and convergence. The paper highlights the importance of lightweight and efficient architectures in medical imaging, emphasizing that they can significantly improve performance without additional pretraining or computational resources. SegFormer3D represents a valuable advancement in 3D medical image segmentation, offering a balance between performance and efficiency that is crucial for practical applications in healthcare.