Subspace-based methods for system identification have attracted much attention during the past few years. This interest is due to the ability of providing accurate state-space models for multivariable linear systems directly from input-output data. The methods have their origin in classical state-space realization theory as developed in the 1960s. The main computational tools are the QR and the singular-value decompositions. Here, an overview of existing subspace-based techniques for system identification is given. The methods are grouped into the classes of realization-based and direct techniques. Similarities between different algorithms are pointed out, and their applicability is commented upon. We also discuss some recent ideas for improving and extending the methods. A simulation example is included for comparing different algorithms. The subspace-based approach is found to perform competitive with respect to prediction-error methods, provided the system is properly excited.