DMPfold tends to produce low-variability models - it produces a small number of distinct conformations, and either the right one is in there or it isn't. This arises from the way the models are generated from distance constraints. Rosetta, by contrast, produces many different conformations which is why 5000 or more models tend to be generated. This arises from Rosetta's method of assembling small fragments, which allows for many more combinations.
You could generate 5000 models with DMPfold. Running with
10 5000 would give you 5000 models as output from the last iteration. Running with
10 500 would give you 500 models from the last iteration but you could also use the 500 models from each of the previous iterations (saved as
ensemble.1.pdb etc.) to get 5000 models overall.
However I wouldn't recommend it, it sounds like a waste of disk space and compute time. Just run with default parameters
3 50 and use the final models saved as
For some numbers on how DMPfold compares to larger numbers of Rosetta models, and how much better DMPfold gets when you use more models, see Table 1 of the paper.