Low-rank sparse (L-S)-reconstruction has been successfully applied to k-t-accelerated applications like cardiac cine imaging. We propose a novel way of k-t rank separation for the reconstruction of non-Cartesian parallel fMRI. Instead of reconstructing the fMRI images separately, the proposed method reconstructs images jointly. This method extracts temporal signal variation information from k-t space directly, thus exactly preserving dynamic information. The results show a higher dynamic signal recovery rate and shorter reconstruction time.