We have carried out a set of explicit solvent molecular dynamics (MD) simulations on two DNA quadruplex (G-DNA) molecules, namely the antiparallel d(G4T4G4)2 dimeric quadruplex with diagonal loops and the parallel-stranded human telomeric monomolecular quadruplex d[AGGG(TTAGGG)3] with three propeller loops. The main purpose of the paper was testing of the capability of the MD simulation technique to describe single-stranded topologies of G-DNA loops, which represent a very challenging task for computational methods. The total amount of conventional and locally enhanced sampling (LES) simulations analyzed in this study exceeds 1.5 μs, while we tested several versions of the AMBER force field (parm99, parmbsc0, and a version with modified glycosidic χ torsion profile) and the CHARMM27 force field. Further, we compared minimal salt and excess salt simulations. Postprocessing MM-PBSA (Molecular Mechanics, Poisson-Boltzmann, Surface Area) free energy calculations are also reported. None of the presently available force fields is accurate enough in describing the G-DNA loops. The imbalance is best seen for the propeller loops, as their experimental structure is lost within a few ns of standard simulations with all force fields. Among them, parmbsc0 provides results that are clearly closest to the experimental target values but still not in full agreement. This confirms that the improvement of the γ torsional profile penalizing the γ trans substates in the parmbsc0 parametrization was a step in the right direction, albeit not sufficient to treat all imbalances. The modified χ parametrization appears to rigidify the studied systems but does not change the ultimate outcome of the present simulations. The structures obtained in simulations with the modified χ profile are predetermined by its combination with either parm99 or parmbsc0. Experimental geometries of diagonal loops of d(G4T4G4)2 are stable in standard simulations on the ∼10 ns time scale but are becoming progressively lost in longer and LES simulations. In addition, the d(G4T4G4)2 quadruplex contains, besides the three genuine binding sites for cations in the channel of its stem, also an ion binding site at each stem-loop junction. This arrangement of five cations in the quadruplex core region is entirely unstable in all 24 simulations that we attempted. Overall, our results confirm that G-DNA loops represent one of the most difficult targets for molecular modeling approaches and should be considered as reference structures in any future studies aiming to develop or tune nucleic acids force fields.