We use 50 Tonga-Fiji events recorded at the broadband TriNet array, southern California, to develop a pure path upper mantle shear velocity model across the Pacific. At the epicentral distances of 70°–95°, multibounce S waves up to S ⁵ are observed, including their triplicated branches, which become particularly clear after stacking. Since these S wave multiples turn at various depths, simultaneously modeling their differential traveltimes and waveforms provides strong constraints on the radial velocity structure. We parameterize the velocity model according to a priori information from the previous oceanic models, so that we can take a grid search approach, to fully investigate possible interdependencies among the model parameters. We construct synthetics with a reflectivity code and study both the SH and SV components. By modeling the whole recordings from events at different depths, with different mechanisms, we are able to separate shallow low-velocity zone (LVZ) features from deeper structure. Our preferred model (PAC06) contains a fast lid (V _sh = 4.78 km s⁻¹, V _sv = 4.58 km s⁻¹) with a thickness of ∼60 km. The underlying LVZ is prominent with the lowest velocities V _sh = 4.34 km s⁻¹, and V _sv = 4.22 km s⁻¹ occurring at a depth of ∼160 km. These velocities are below the estimates of solid-state LVZ, suggesting the presence of partial melt. The anisotropy (V _sv < V _sh ) of PAC06 extends to ∼300 km depth, which is constrained by S ⁵ turning at this depth. Besides the 406 km and 651 km discontinuities, PAC06 also has a small (∼1%) velocity jump at ∼516 km. We consider these main features of PAC06 to be well determined, since PAC06 explains a large data set from various events. Therefore it is ideally suited for comparing with mineralogical models.