Link adaptation allows a communication system to adapt its transmission modes according to channel conditions. Although a maximum likelihood (ML) receiver for multiple-input multiple-output (MIMO) systems provides optimal performance, estimating its link performance has been a difficult problem. In this paper, we propose a new link performance abstraction technique for MIMO orthogonal frequency-division multiplexing systems with the ML receiver. The performance of ML detection (MLD) is estimated by employing capacity bounds of two simple linear receivers. Then, we give a simple parametrization to compute the desired per-stream signal-to-noise ratio (SNR) values, which can be applied for both vertically and horizontally coded MIMO systems. Based on the derived per-stream SNR estimates, the block error rate is obtained using the received-bit information rate metrics. We also examine the effect of imperfect channel estimation as well as spatial correlations among antennas. Finally, extensive simulation results show that the proposed method provides superior estimation accuracy in the MIMO-MLD link evaluation with very low computational complexity.