This paper presents a variational network (VN) for the reconstruction of accelerated MRI data. The VN combines the mathematical structure of variational models with deep learning to enable fast and high-quality reconstruction of clinical multi-coil MRI data. The VN is trained on a clinical knee imaging protocol and evaluated on different acceleration factors and sampling patterns. The results show that the VN outperforms standard reconstruction algorithms in terms of image quality and residual artifacts. The VN preserves the natural appearance of MR images and pathologies not included in the training data. It also offers high computational performance, with a reconstruction time of 193 ms on a single graphics card. The VN is trained offline and can be applied online to previously unseen data without parameter tuning. The VN is designed to learn a complete reconstruction procedure for complex-valued multichannel MR data, including all free parameters. The VN is trained on a complete clinical protocol for musculoskeletal imaging and evaluated for different acceleration factors and sampling patterns. The VN is able to preserve unique pathologies not included in the training data. The VN is implemented in C++/CUDA and provides Python and Matlab interfaces for testing. The VN is trained on a system equipped with an Intel Xeon E5-2698 CPU and a single Nvidia Tesla M40 GPU. The VN is compared to the linear PI reconstruction method CG SENSE and a combined PI-CS non-linear reconstruction method based on Total Generalized Variation (TGV). The VN outperforms both methods in terms of image quality and reconstruction speed. The VN is able to suppress undersampling artifacts and provide sharper and more natural-looking images. The VN is able to handle complex-valued multi-coil data and is designed to be efficient and flexible. The VN is able to adapt to different types of image features and artifact properties. The VN is able to capture more efficiently the characteristic backfolding artifacts of Cartesian undersampled data. The VN is able to suppress artifacts with PI-CS TGV by choosing appropriate regularization parameters. The VN is able to handle a wide range of pathologies and offers high reconstruction speed, which is substantial for integration into clinical workflow. The VN is able to achieve high-quality reconstructions even with low SNR. The VN is able to handle different sampling patterns and is designed to be efficient and flexible. The VN is able to adapt to different types of image features and artifact properties. The VN is able to capture more efficiently the characteristic backfolding artifacts of Cartesian undersampled data. The VN is able to suppress artifacts with PI-CS TGV by choosing appropriate regularization parameters. The VN is able to handle a wide range of pathologies and offers high reconstruction speed, which is substantial for integration into clinical workflow. The VN is able to achieve high-quality reconstructions even with low SNR.This paper presents a variational network (VN) for the reconstruction of accelerated MRI data. The VN combines the mathematical structure of variational models with deep learning to enable fast and high-quality reconstruction of clinical multi-coil MRI data. The VN is trained on a clinical knee imaging protocol and evaluated on different acceleration factors and sampling patterns. The results show that the VN outperforms standard reconstruction algorithms in terms of image quality and residual artifacts. The VN preserves the natural appearance of MR images and pathologies not included in the training data. It also offers high computational performance, with a reconstruction time of 193 ms on a single graphics card. The VN is trained offline and can be applied online to previously unseen data without parameter tuning. The VN is designed to learn a complete reconstruction procedure for complex-valued multichannel MR data, including all free parameters. The VN is trained on a complete clinical protocol for musculoskeletal imaging and evaluated for different acceleration factors and sampling patterns. The VN is able to preserve unique pathologies not included in the training data. The VN is implemented in C++/CUDA and provides Python and Matlab interfaces for testing. The VN is trained on a system equipped with an Intel Xeon E5-2698 CPU and a single Nvidia Tesla M40 GPU. The VN is compared to the linear PI reconstruction method CG SENSE and a combined PI-CS non-linear reconstruction method based on Total Generalized Variation (TGV). The VN outperforms both methods in terms of image quality and reconstruction speed. The VN is able to suppress undersampling artifacts and provide sharper and more natural-looking images. The VN is able to handle complex-valued multi-coil data and is designed to be efficient and flexible. The VN is able to adapt to different types of image features and artifact properties. The VN is able to capture more efficiently the characteristic backfolding artifacts of Cartesian undersampled data. The VN is able to suppress artifacts with PI-CS TGV by choosing appropriate regularization parameters. The VN is able to handle a wide range of pathologies and offers high reconstruction speed, which is substantial for integration into clinical workflow. The VN is able to achieve high-quality reconstructions even with low SNR. The VN is able to handle different sampling patterns and is designed to be efficient and flexible. The VN is able to adapt to different types of image features and artifact properties. The VN is able to capture more efficiently the characteristic backfolding artifacts of Cartesian undersampled data. The VN is able to suppress artifacts with PI-CS TGV by choosing appropriate regularization parameters. The VN is able to handle a wide range of pathologies and offers high reconstruction speed, which is substantial for integration into clinical workflow. The VN is able to achieve high-quality reconstructions even with low SNR.