We modeled Pnl waveforms from two moderate-sized normal-fault earthquakes in Zambia to determine an upper mantle P wave velocity model for southern Africa. The event parameters are May 15, 1968, mb=5.7, depth=28 km; and December 12, 1968, mb=5.9, depth=6 km. Teleseismic body wave inversions by previous researchers constrain the focal parameters for these events. We compute synthetic seismograms for six upper mantle velocity models using a wave number integration algorithm until an acceptable fit to the data is obtained. These models have different mantle P wave velocity gradients and low-velocity zone (LVZ) depths. Quality of fit is measured primarily by the Pn/PL amplitude ratio. Source-station geometry allows for the independent sampling of the upper mantle beneath the Kaapvaal-Zimbabwe craton and mobile belt structural provinces. Synthetics from a constant velocity mantle model do not contain prominent precursor arrivals seen in the data; these are interpreted as P waves turning in the upper mantle. These synthetics also produce a too low Pn/PL amplitude ratio. Synthetics for a model with a mantle P wave velocity gradient of 0.0033 s-1 and a LVZ no shallower than 170 km fit the cratonic path data adequately. A slightly lower gradient, with a LVZ no shallower than 150 km, is indicated for the mobile belt regions. All the data preclude a mantle lid velocity gradient as high as 0.0066 s-1. Different velocity gradient between the two provinces implies different thermal gradients, which supports the hypothesis that a deep, cool lithospheric root exists beneath the Kaapvaal-Zimbabwe craton. Cratonic thermal gradients derived from our velocity gradients agree closely with those derived from thermal modeling by Ballard and Pollack <1987>. ¿ American Geophysical Union 1990