Most inversions of the long-wavelength geoid in conjunction with the seismic tomographic information have so far been carried out under the assumption of either purely whole mantle or perfectly layered circulation. Moreover, modeling the lithosphere as a spherical shell with a uniform low viscosity was found to yield the best fit to the observed geoid. We have tested whether a good prediction of the geoid can also be achieved by including two constraints: a semipermeable behavior of the 660-km discontinuity and surface plate velocities equal to the observed ones. The mass transfer between upper and lower mantle has been changed by imposing a surface density anomaly at a depth of 660 km, which is proportional to the mass anomaly needed to achieve perfectly layered circulation. On the top of the mantle we assume a stiff lithosphere that moves with a velocity corresponding to the observed plate motion. The viscosity only varies with depth. Considering a simple three-layer viscosity structure and changing the permeability of the 660-km interface, we have obtained a satisfactory variance reduction of the geoid data (~75% for degrees 2--12). The best fit to the geoid is obtained if the mass transfer across the 660-km boundary is reduced to one third in comparison with the purely whole mantle model. The best fitting viscosity profile is characterized by a clearly defined asthenosphere and a viscosity increase by at least 2 orders of magnitude in the lower mantle. The amplitudes of dynamic topography predicted by our model are remarkably small (~100 m), thus fully compatible with the observation. ¿ 1999 American Geophysical Union