Looking on down the wiki, it seems like the "free-forecast" is associated with the deterministic model, while the "cycled" is associated with the Ensemble Kalman Filter. That is, the free-forecast is not constrained by data assimilation algorithms- it can just start from a restart file and go. A cycled run, however, needs multiple runs, as well as data.

This is all speculation based on naming conventions. If you really want an answer, perhaps asking the code developer?