In this paper, we propose an optimized tensor model for direction of departure (DOD) and direction of arrival (DOA) estimation with coprime multiple input multiple output (MIMO) radar. We first develop the Toeplitz matrix using the consecutive lags in the sum-difference coarray after vectorizing the array covariance matrix. We then construct a generalized three-way tensor model and derive the optimized tensor by maximizing the number of targets. Then, the existing COMFAC technique is exploited for angle estimation. We show that the proposed method can identify more targets and achieve better performance by enforcing the three-way structure as compared to the subspace-based algorithms. Numerical simulations are provided to validate the theoretical analysis.