In order to achieve real-time image rendering, optoacoustic tomography reconstructions are commonly done with back-projection algorithms due to their simplicity and low computational complexity. However, model-based algorithms have been shown to attain more accurate reconstruction performance due to their ability to model arbitrary detection geometries, transducer shapes and other experimental factors. The high computational complexity of the model-based schemes makes it challenging to be implemented for real time inversion. Herein, we introduce a novel discretization method for model-based optoacoustic tomography that enables its efficient parallel implementation on graphics processing units with extremely low memory overhead. We demonstrate that, when employing a tomographic scanner with 256 detectors, the new method achieves model-based optoacoustic inversion at 20 frames per second for a 200 × 200 image grid.