HMMs rely on primary sequence consensus, and though they're pretty powerful models, you shouldn't expect HMMs to recognize (or align) very distant structural homologues.
One drawback of the current package is that it doesn't make very good models of small numbers of example sequences. I have had best luck when I have more than 50-100 example sequences. If you have only three or four, expect trouble. We are working on principled methods to overcome this difficulty. In the meantime, check out the ad hoc -P options to hmmb and hmmt in the man pages. This option lets you incorporate PAM matrix information into a model in a semi-principled way.