It very much depends on what type of defects you are considering. My experience is that at least if you are looking strictly at carbon, and the defects are geometrical, meaning some atoms missing, or you are tailoring the shape of a nanoribbon, adding Stone-Wales defects to nanotubes, etc, then even very simple tight-binding models can give a qualitative picture that is fairly accurate. But if the question is more about various chemical species being absorbed, or there is a very specific dependence on bond lengths or out-of-plane effects such as warping or buckling, then probably only DFT will capture the effects.
The max interaction range can be changed a bit, the default is a bit low, but increasing it also can make the Huckel models for instance about as slow as DFT. It's more important to have the correct dependence of the matrix elements with bond lengths and angles, than necessarily have 4th, 5th, etc neighbor interactions.
In general, I do not recommend the CP2K basis sets without careful testing - simply because I haven't tested them myself very much, and when I did I wasn't happy about the results. I would rather say, if you are looking for very simple things like graphene ribbons where the shape changes a lot, try the simple pi model by Hancock, else for a bit more complex stuff the Cerda model for graphite (also works nice for graphene - but remember to use Hoffmann Hydrogen!). With some testing, perhaps the sp3d5s* models of Jancu, Bassani can be useful.
For DFTB (whether Hotbit, CP2K, or dftb.org parameters) I would first test the parameters against known band structures, like graphene, nanoribbons, etc, where a comparison with DFT is cheap. If you get decent results, then you can use it for transport
