It is an issue that multivariable systems with high dimensions have many parameters, resulting in heavy computational costs in identification methods. Therefore, a coupled stochastic gradient algorithm is derived for multivariable systems based on the coupling identification concept. The identification model is decomposed into several single-output systems, and the parameter estimates are coupled during the subsystem identification by using the gradient search. The convergence properties are analyzed by using the martingale convergence theorem. Compared with the recursive least squares algorithm and the coupled least squares algorithm, the proposed algorithm has less computational load. The convergence rate can be improved by introducing a forgetting factor. Performance analysis verifies that the proposed algorithm converges. The simulation results show the effectiveness of the proposed algorithm.