We propose a cumulant-based algorithm for multiple near-field source localization based on exact source-sensor spatial geometry, where no simplification, such as Fresnel approximation of spatial phase-delay and inconsideration of propagation attenuation is made. We define five cumulant matrices, by which a set of coarse angle-range estimates plus a set of ambiguous angle-range estimates are derived. Fine and unambiguous estimation is finally obtained by a simple refinement step. We define the algorithm's name as Cumulant Algorithm For Exact Model (CAFEM). Compared with previous algorithms, CAFEM has the following advantages: (1) it is simple, analytical, and search-free, (2) it is not necessary to use uniformly spaced linear arrays, and (3) it is applicable to array sensors with unknown amplitude uncertainties and can accommodate any arbitrarily unknown propagation loss.