Deep earthquakes of up to moderate size have been analyzed by a long-period waveform inversion of Global Digital Seismographic Network data where the use of a point source model proved to be adequate, but the use of a standard earth model was inadequate without modification. Modifications included calculations in an earth model based on PREM but with different source and receiver structures and the introduction of time delays which were determined together with the moment tensor components of the source by an iterative correlation method. The moment tensor components are in accordance with the pattern of stress release in other work, but results for the seismic moment differ somewhat and there is no compelling evidence for deviations from a double couple mechanism. The time delays were inverted to find the source centroid depths and time shifts and the station residuals. There is no baseline effect in the times of source duration which could be explained by a simple circular source model, but there is such an effect in the station residuals together with a distance trend especially for S. Part of the baselines may be an effect of bias in origin times, but the pattern of S residuals in particular suggests also possible modifications of the standard model.