The paper presents a novel approach, termed Variational Network (VN), for efficient reconstruction of complex multi-coil MRI data. The VN combines variational models and deep learning, formulating image reconstruction as a variational model embedded in an unrolled gradient descent scheme. The parameters of this model, including filter kernels, activation functions, and data term weights, are learned during an offline training procedure. The learned model can then be applied online to previously unseen data, providing high-quality reconstructions with reduced artifacts and improved image quality. The VN is evaluated on a clinical knee imaging protocol, demonstrating superior performance compared to standard reconstruction algorithms in terms of image quality and residual artifacts for various acceleration factors and sampling patterns. The VN preserves the natural appearance of MR images and pathologies not included in the training data set. Additionally, the VN offers high computational performance, with a reconstruction time of 193 ms on a single graphics card, making it suitable for integration into clinical workflow. The key contributions of the work include the development of a variational network architecture that combines variational methods and deep learning, the theoretical foundation for understanding the learned model, and the demonstration of the VN's effectiveness in practical clinical applications. The VN approach provides a flexible and efficient solution for accelerated MRI reconstruction, addressing the challenges of complex multi-coil data and undersampled acquisitions.The paper presents a novel approach, termed Variational Network (VN), for efficient reconstruction of complex multi-coil MRI data. The VN combines variational models and deep learning, formulating image reconstruction as a variational model embedded in an unrolled gradient descent scheme. The parameters of this model, including filter kernels, activation functions, and data term weights, are learned during an offline training procedure. The learned model can then be applied online to previously unseen data, providing high-quality reconstructions with reduced artifacts and improved image quality. The VN is evaluated on a clinical knee imaging protocol, demonstrating superior performance compared to standard reconstruction algorithms in terms of image quality and residual artifacts for various acceleration factors and sampling patterns. The VN preserves the natural appearance of MR images and pathologies not included in the training data set. Additionally, the VN offers high computational performance, with a reconstruction time of 193 ms on a single graphics card, making it suitable for integration into clinical workflow. The key contributions of the work include the development of a variational network architecture that combines variational methods and deep learning, the theoretical foundation for understanding the learned model, and the demonstration of the VN's effectiveness in practical clinical applications. The VN approach provides a flexible and efficient solution for accelerated MRI reconstruction, addressing the challenges of complex multi-coil data and undersampled acquisitions.