The recently introduced notion of peak arrivals [Athanassoulis and Skarsoulis, J. Acoust. Soc. Am. 97, 3575–3588 (1995)], defined as the significant local maxima of the arrival pattern, is studied here as a modeling basis for performing ocean tomography. Peak arrivals constitute direct theoretical counterparts of experimentally observed peaks, and offer a complete modeling of experimental observables, even in cases where ray or modal arrivals cannot be resolved. The coefficients of the resulting peak‐inversion system, relating travel‐time with sound‐speed perturbations, are explicitly calculated in the case of range‐independent environments using normal‐mode theory. To apply the peak‐inversion scheme to tomography the peak identification and tracking problem is examined from a statistical viewpoint; maximum‐likelihood and least‐square solutions are derived and discussed. The particular approach adopted treats the identification and tracking problem in close relation to the inversion procedure; all possibilities of associating observed peaks with background arrivals are examined via trial inversions, and the best peak identification is selected with respect to a least‐square criterion. The feasibility of peak tomography is subsequently demonstrated using first synthetic data and then measured data from the THETIS‐I experiment. In the synthetic case the performance of the overall scheme is found to be satisfactory both with noise‐free and noisy data. Furthermore, the identification, tracking, and inversion results using experimental acoustic data from THETIS‐I are in good agreement with independent field observations.