No, I don't think the model should be decimated for this particular purpose. Turning ObjReader's format to binary (some day) will solve this problem automatically. In the meantime, the models should be left intact regardless of extra time IMHO. There are not so many models whose load time is downright unbearable anyway.
I have been thinking for long to use a kind of bload/bsave to convert from .obj into .orb, using a compound format.
That is very high on my list, leaving FBO post-processing to your sole discretion
