We present Matrioshka orthogonal matching pursuit (OMP), a method consisting of two nested OMPs for separating seismic sources at an early stage of the signal processing chain. Matrioshka OMP is based on models of sensor signals that place nonrestrictive assumptions on the seismic survey using simultaneous sources. Our seismic event model is based on the spatial coherence of signals, which results in a straight or slightly curved feature in the trace representation of the data with a specific wavelet, whose magnitude can linearly vary according to the offset. We demonstrate the effectiveness of the approach on synthetic and real data.