This is not a easy question, but results from literature indicate that hybrid functionals such as B3LYP (in a BS scheme) usually are more accurate than GGAs for the prediction of exchange coupling constants. I do not know which the best choice in yours particular case, but I suggest you to start with B3LYP and/or B3P86 functionals. 

There are some review articles about BS approach inside a DFT formalism. For instance:

Photosynth Res (2009) 102:443–453

doi:10.1016/j.ccr.2008.05.014

Cheers

Hello Michael,

If you are interested in use an hybrid functional, I suggest to use B3PW91 or taking in account that it must be localized in a protein, a Minesota functional could help depending on cluster and interactions. About B3LYP I suggest to don't use it because can give bad results, and for metals, B3LYP can't explain the homogeneous electron gas an introduces a important errors.

About basis set I think that Dunning are a very good option but if you need more time-control a Pople ones can offer more combination (at less accuracy). Also pseudo-potential could be a very good option.

I add two link if you are interested in B3PW91 instead of popular B3LYP.

I hope it helps you,

Joaquim Rius

https://www.researchgate.net/post/What_basis_and_functional_set_would_be_suitable_for_ab_initio_calculations_of_copper_and_zinc_ions_an_organic_compound_and_water

https://www.researchgate.net/post/Why_B3LYP_was_not_re-optimized

It's hard to justify the use of B3LYP today, since there are so many better functionals, giving lower errors in different benchmarks.

When doing organometallic studies I usually stick to M06 series of functionals, with the D3 dispersion correction from Grimme. This series of functionals is known to give great results in various benchmarks - both with small standard molecules (http://comp.chem.umn.edu/info/docs/773.pdf) and in systems with metals (http://scitation.aip.org/content/aip/journal/jcp/132/16/10.1063/1.3382342). They're also implemented in NWChem (although I'm not sure about the D3 correction). I usually do the geometry optimization (+vibrations) at the double-zeta level and then single-point at the triple-zeta level: but recently one of the reviewers of my work suggested the triple-zeta already at the optimization level. I don't find it necessary, since it basically doubles the computational time, while not really improving the results a lot.

A recent publication by Harris & Szilagyi

http://www.ncbi.nlm.nih.gov/pubmed/24458434

J Comput Chem. 2014 Mar 15;35(7):540-52. doi: 10.1002/jcc.23518comparing various functionals and basis sets suggests the B(5%HF)P86 functional, (5% Hartee -Fock exchange) together with the Def2-TZVP basis set for iron-sulfur clusters. Whereas the basis set is ready to use, I'm uncertain about the correct specification of this particular DFT functional in NWChem.

From general considerations PBE0 should be better than B3LYP as the latter is parametrized only on p-elements sets.