In this paper we present a new method to compare the topological structure of two continuous random network (CRN) models in a quantitative manner. The comparison serves to assess the sampling properties of the molecular modeling program TUMBLEWEED. This program generates spherical CRN clusters and has been introduced in paper I of this series. The structural diversity of the models constructed with this program is found to be satisfactory. Further, the models are very different from the α-cristobalite and α-quartz structures of GeO 2. In addition, we investigate three different techniques for the refinement of the clusters presented in paper I: the conjugate gradient (CG) method, the Monte Carlo (MC) method, and the Reverse Monte Carlo (RMC) method. The former two procedures minimize the potential energy, V, whereas the latter minimizes the deviation, R χ , of the model neutron total correlation function, T( r), from corresponding experimental data. It is found that neither technique gives satisfactory results by itself, since the minimization of R χ leads to an increase of V (including the violation of geometric constraints) and vice versa. By combining the MC and RMC methods, i.e. by simultaneous minimization of V and R χ , we obtain models with realistic geometric properties and an improved fit to the experimental data ( R χ =0.018±0.003 in the range 0–1 nm). This further supports the validity of the CRN approach to the structure of glassy GeO 2. New models with a non-uniform torsion angle distribution have been constructed. It turns out that such models cannot be excluded as possible structures (final R χ =0.020±0.006).