Mantle convection models based on geophysical constraints have provided us with a basic understanding of the forces driving and resisting plate motions on Earth. However, existing studies computing the balance of underlying forces are contradicting, and the impact of plate boundary geometry on surface deformation remains unknown. We address these issues by developing global instantaneous 3-D mantle convection models with a heterogeneous density and viscosity distribution and weak plate boundaries prescribed using different geometries. We find that the plate boundary geometry of the Global Earthquake Model (GEM, Pagani et al., 2018), featuring open plate boundaries with discrete lithospheric-depth weak zones in the oceans and distributed crustal faults within continents, achieves the best fit to the observed GPS data with a directional correlation of 95.1% and a global point-wise velocity residual of 1.87 cm/year. A good fit also requires plate boundaries being 3 to 4 orders of magnitude weaker than the surrounding lithosphere and low asthenospheric viscosities between 5e17 and 5e18 Pa s. Models without asthenospheric and lower mantle heterogeneities retain on average 30% and 70% of the plate speeds, respectively. Our results show that Earth’s plate boundaries are not uniform and better described by more discrete plate boundaries within the oceans and distributed faults within continents. Furthermore, they emphasize the impact of plate boundary geometry on the direction and speed of plate motions and reaffirm the importance of slab pull in the uppermost mantle as a major plate driving force.