A previously proposed nonlinear inverse reconstruction for autocalibrated parallel imaging simultaneously estimates coil sensitivities and image content. This work exploits this property for real-time MRI, where coil sensitivities need to be dynamically adapted to the conditions generated by moving objects. The development comprises (i) an extension of the nonlinear inverse algorithm to non-Cartesian k-space encodings, (ii) its implementation on a graphical processing unit to reduce reconstruction times, and (iii) the use of a convolution-based iteration, which considerably simplifies the graphical processing unit implementation compared to a gridding technique. The method is validated for real-time MRI of the human heart at 3 T using radio frequency-spoiled radial FLASH (pulse repetition time/echo time = 2.0/1.3 ms, flip angle 8°). The results demonstrate artifact-free reconstructions from only 65–85 spokes, with 256 oversampled data points. Acquisition times of 130–170 ms resulted in 29–38 frames per second for sliding window reconstructions (factor 5). While offline reconstructions required 1–2 sec, real-time applications with modified parameters and slightly lower image quality were achieved within 90 ms per graphical processing unit. Magn Reson Med 63:1456–1462, 2010. © 2010 Wiley-Liss, Inc.