When you do homology modelling, you basically replace the residues on the template with those of the sequence you are modelling. Each modelled residue is inserted in a complex environment and needs to 'adapt' to it This is what energy minimization is there for. Imagine that you have a tryptophan residue being modelled where a glycine was in the template. For sure you'll have some steric clashes due to the larger size of the tryptophan, the side chain interactions in the vicinity need to be readjusted, etc. Energy minimization simply brings the system to a point of lower potential energy, thus more stable, and is used to clear up these issues.
Energy minimization will rarely 'ruin' your model, it will usually clear up some bad stereochemistry, but it won't do miracles.