We investigated four key aspects of forward models for distributed solutions to the MEG inverse problem: 1) the nature of the cortical mesh constraining sources (derived from an individual's MRI, or inverse-normalised from a template mesh); 2) the use of single-sphere, overlapping spheres, or Boundary Element Model (BEM) head-models; 3) the density of the cortical mesh (3000 vs. 7000 vertices); and 4) whether source orientations were constrained to be normal to that mesh. These were compared within the context of two types of spatial prior on the sources: a single prior corresponding to a standard L2-minimum-norm (MNM) inversion, or multiple sparse priors (MSP). The resulting generative models were compared using a free-energy approximation to the Bayesian model-evidence after fitting multiple epochs of responses to faces or scrambled faces. Statistical tests of the free-energy, across nine participants, showed clear superiority of MSP over MNM models; with the former reconstructing deeper sources. Furthermore, there was 1) no evidence that an individually-defined cortical mesh was superior to an inverse-normalised canonical mesh, but 2) clear evidence that a BEM was superior to spherical head-models, provided individually-defined inner skull and scalp meshes were used. Finally, for MSP models, there was evidence that the combination of 3) higher density cortical meshes and 4) dipoles constrained to be normal to the mesh was superior to lower-density or freely-oriented sources (in contrast to the MNM models, in which free-orientation was optimal). These results have practical implications for MEG source reconstruction, particularly in the context of group studies.