U-Mamba is a novel network designed for biomedical image segmentation that addresses the limitations of CNNs and Transformers in capturing long-range dependencies. Inspired by State Space Sequence Models (SSMs), U-Mamba integrates the local feature extraction of CNNs with the long-range dependency modeling capabilities of SSMs. It features a self-configuring mechanism, allowing it to adapt to various datasets without manual intervention. The network uses a hybrid CNN-SSM architecture, combining Residual blocks with Mamba blocks to capture both local and long-range features. U-Mamba outperforms state-of-the-art CNN-based and Transformer-based segmentation networks across four diverse tasks, including 3D abdominal organ segmentation, endoscopy instrument segmentation, and microscopy cell segmentation. It achieves superior performance with linear scaling in feature size and efficient computation. The results show that U-Mamba has fewer segmentation outliers and better global context understanding, particularly in handling objects with heterogeneous appearances. The network is implemented within the nnU-Net framework and trained on various medical image datasets. U-Mamba demonstrates strong performance in both 3D and 2D segmentation tasks, achieving high Dice Similarity Coefficient (DSC) and Normalized Surface Distance (NSD) scores. The results indicate that U-Mamba is a promising candidate for future biomedical image segmentation networks due to its efficient long-range dependency modeling capabilities. The code, models, and data are publicly available for further research and development.U-Mamba is a novel network designed for biomedical image segmentation that addresses the limitations of CNNs and Transformers in capturing long-range dependencies. Inspired by State Space Sequence Models (SSMs), U-Mamba integrates the local feature extraction of CNNs with the long-range dependency modeling capabilities of SSMs. It features a self-configuring mechanism, allowing it to adapt to various datasets without manual intervention. The network uses a hybrid CNN-SSM architecture, combining Residual blocks with Mamba blocks to capture both local and long-range features. U-Mamba outperforms state-of-the-art CNN-based and Transformer-based segmentation networks across four diverse tasks, including 3D abdominal organ segmentation, endoscopy instrument segmentation, and microscopy cell segmentation. It achieves superior performance with linear scaling in feature size and efficient computation. The results show that U-Mamba has fewer segmentation outliers and better global context understanding, particularly in handling objects with heterogeneous appearances. The network is implemented within the nnU-Net framework and trained on various medical image datasets. U-Mamba demonstrates strong performance in both 3D and 2D segmentation tasks, achieving high Dice Similarity Coefficient (DSC) and Normalized Surface Distance (NSD) scores. The results indicate that U-Mamba is a promising candidate for future biomedical image segmentation networks due to its efficient long-range dependency modeling capabilities. The code, models, and data are publicly available for further research and development.