Semi-empirical methods are parameterized on the basis of selected properties of a selected set of molecules. A reasonable performance can be expected for related compounds, but for problems that are not covered by the "training set" the reliability is limited. In this respect ab initio methods, even low-level ones, are more widely applicable.
For molecular structure and heats of formation of closed-shell molecules MNDO, AM1 and PM3 are quite good. Practical experience has shown that for some particular problems one of the three performs markedly better than the others, but in general the most recent methods AM1 and PM3 are preferred. PM3 is parameterized for a greater number of elements, but sometimes the parameters are based upon a very small set of data. The mean absolute errors of the bond lengths between heavy atoms are reported to be 0.036 Å for PM3 and slightly greater for AM1 and MNDO. The error in bond angles is 3 - 4 degrees . Even low-level ab initio calculations are usually better, but very much more time-consuming.
Trends in vibrational frequencies in related systems are usually reproduced by semi-empirical calculations, but the errors are not as nicely systematic (and therefore correctable) as in ab initio calculations. When energy is concerned the results are not completely satisfactory, in spite of the fact that the methods are parameterized to reproduce heats of formation.
 Spartan User's Guide, version 3.0, Wavefunction, Inc., 1993.