SUMMARY The structure of the Earth’s upper mantle near convergent plate margins, such as along the Nubia‐Eurasia collision zone in the Mediterranean, involves strong seismic wave speed contrasts associated with subducting lithospheric slabs and opening backarc basins. In this environment, seismic wave propagation is strongly influenced by heterogeneity, and requires appropriate modelling practice. Although accurate numerical methods are often used to model seismic traveltimes in the crust, only approximate techniques have been used for the mantle, on the assumption that speed contrasts are weaker. We devise, optimize, and test a method aimed at recovering strongly heterogeneous mantle structures using a finite-difference scheme to calculate first-arrival traveltimes and trace seismic rays with high accuracy even in the presence of strong gradients. We adapt this forward scheme—successfully used in local-scale tomography—to spherical geometry through source-specific Earth flattening approximations, and we split calculations in meshes with different step size to model optimally the crust (with a 2 km step) and the mantle (6 km step). We then use an iterative non-linear inversion approach, starting from a simple 1-D prior model. We test the ability of this procedure to reconstruct sample structures, devised to be illustrative for the Mediterranean region, using synthetic data calculated on the real distribution of sources and stations reported by the Bulletins of the International Seismological Centre (ISC). Besides regular checkerboard patterns, we also reconstruct a more representative model. Different strategies are used and compared in linear and non-linear inversion. We find that a linear approach, by which rays are only traced once in the background model, may result in an illusory fit to data. Realistic upper-mantle structures strongly deflect seismic rays, and correct paths can only be found after a few iterations. Although linear inversion seems able to identify the main features quite well, we verify how non-linear inversion and 3-D ray tracing significantly improve the results, especially when we attempt to reconstruct a realistic structure. We also apply the finite-difference, non-linear, traveltime tomography to data from the ISC to retrieve upper-mantle structure in the Central Mediterranean. We verify that the non-linear inversion is able to reveal sharpened velocity contrasts and thinner bodies than linear inversion. Clear differentiation found in the nonlinear result, between signatures of northern and southern Dinarides—showing lithosphere subducting only beneath the southern sector—is more coherent with the regional geodynamic framework. Such improvements due to non-linear mantle tomography may contribute to the general picture of slab detachment and small-scale mantle convection in the Mediterranean region, and therefore, significantly impact on geodynamic implications of resulting models.