A thorough ab initio study of the Mn(2) dimer in its lowest electronic states that correlate to the ground Mn((6)S)+Mn((6)S) dissociation limit is reported. Performance of multireference methods is examined in calculations of the fully spin-polarized S=5((11) summation operator(+) (u)) state against the recent accurate single-reference coupled cluster CCSD(T) results [A. A. Buchachenko, Chem. Phys. Lett. 459, 73 (2008)]. The detailed comparison reveals a serious disagreement between the multireference configuration interaction (MRCI) and related nonperturbative results on the one hand and the complete active space perturbation theory (CASPT) calculations on the other. A striking difference found in the CASPT results of the second and third orders indicates poor perturbation expansion convergence. It is shown that a similar problem has affected most of the previous calculations performed using CASPT2 and similar perturbative approximations. The composition of the active space in the reference multiconfigurational self-consistent field calculations, the core correlation contribution, and basis set saturation effects are also analyzed. The lower spin states, S=0-4, are investigated using the MRCI method. The results indicate a similar dispersion binding for all the spin states within the manifold related to the closed 4s shells, which appears to screen and suppress the spin coupling between the half-filled 3d atomic shells. On this premise, the full set of model potentials is built by combining the accurate reference CCSD(T) interaction potential for S=5 and the MRCI spin-exchange energies for the S<5 states. This approach leads to the value of 550 cm(-1) as a lower bound for the (1) summation (+) (g) ground-state dissociation energy. The spin-exchange energies themselves are found to comply with the simple Heisenberg model. The effective spin-coupling parameter J is estimated as -3.9 cm(-1), a value roughly 2.5 times smaller in magnitude than those measured in the inert gas cryogenic matrices. Compressing of the Mn(2) dimer in the matrix cage is suggested as the prime cause of this disagreement.