Magnetoencephalography (MEG) is used in clinical and fundamental studies of brain functions, primarily for the excellent temporal resolution it provides. The spatial resolution is often assumed to be poor, because of the ill-posed nature of MEG source modeling. However, the question of spatial resolution in MEG has seldom been studied in quantitative detail. Here we use the well-known retinotopic organization of the primary visual cortex (V1) as a benchmark for estimating the spatial resolution of MEG source imaging. Using a standard visual stimulation paradigm in human subjects, we find that individual MEG sources exhibit well-delineated visual receptive fields that collectively follow the known mapping of the retinal surface onto the cortex. Based on the size of these receptive fields and the variability of the signal, we are able to resolve MEG signals separated by approximately 7 mm in smooth regions of cortex and less than 1 mm for signals near curved gyri. The maximum resolution is thus comparable to that of the spacing of hypercolumns in human visual cortex. Overall, our results suggest that the spatial resolution of MEG can approach or in some cases exceed that of fMRI.