We use global positioning system (GPS) data to study the rupture mechanism of the 1994 Northridge earthquake in southern California. We include data from 62 observation sites, of which two (Palos Verdes and Jet Propulsion Lab) are permanent GPS geodetic array (PGGA) sites. We use a grid-search scheme to study the range of single- and dual-plane, uniform and varied slip models consistent with the data. We find that in order to fit the geodetic data with a fault model whose primary fault patch is confined to a plane through the aftershocks, a secondary fault plane is required above the primary fault plane. The moment release of the secondary fault can be as large as 1.9 × 1018 N-m, 14% of the moment release of the primary fault. This result implies significant deformation in the shallow crust associated with the mainshock. Our preferred model has a 14 × 14 array of dislocation patches on a plane through the main aftershock cluster and a 5 × 6 array of patches in the hanging wall west of the epicenter. We estimate the displacements on the patches by linear inversion with a first-order smoothness constraint. The estimated displacements on the main fault for this model are confined to a simple region between depths of 5 and 18 km, in the interior of the modeled fault surface. The mainshock lies at the bottom of the aftershock zone, near which about 1-m slip is shown on our modeled fault surface. The maximum slip on the fault surface is about 2.2 m, located at 34.28° N, 118.55° W, and 12.4 km at depth. The seismic moment release estimate of 1.34 ± 0.15 × 1019 N-m on the main fault at the 95% confidence is consistent with the estimate from strong-motion studies.