I am noticing that Modeller does not do a good job when it comes to
sequence alignment of short peptides or DNA sequences.
You're not required to use Modeller to make modeling alignments - as
long as you can make alignment files in the PIR format Modeller wants,
you can use any package (including ClustalW).
That said, by default Modeller's alignment methods use global alignment
and so try to align the termini, since this is typically what you want
for modeling. You can adjust the overhang parameter to discourage this
behavior, or use local alignment.