The combined use of modal and balanced truncations methods is proposed for model order reductions. To efficiently combine these methods, a stopping criterion based on spectral energy concepts is also proposed. This criterion was implemented into the code of the widely known subspace accelerated dominant pole algorithm (SADPA), designed to compute a set of dominant poles and associated residues of transfer functions from large‐scale, sparse, linear descriptor systems. The resulting enhanced SADPA code automatically stops once the computed set of dominant poles and associated residues is sufficient to build a modal reduced order model (ROM) whose energy content approaches that of the complete model within a specified tolerance and considering a frequency window of interest. The number of dominant poles in this set is much smaller than the number of poles of the full system model. Hence, their state‐space realisation usually has a small enough dimension for the efficient application of the square root balanced truncation method. This new method, named hybrid modal‐balanced truncation, produces ROMs whose order are much smaller than that of the modal ROMs and, most importantly, can also be applied to unstable models.