In this paper, a novel iterative MIMO subspace detection method is proposed to approach the maximum-likelihood (ML) receiver performance while keeping the computational complexity scalable for MIMO systems with large number of transmit and receive antennas. By dividing the signal detection problem for a large MIMO system into subspaces with smaller size, the complexity of applying a per-subspace ML receiver is much lower while near ML performance may still be achieved in combination with an iterative parallel interference cancellation unit. It is shown that, for equal number of transmit and receive antennas, the complexity of the proposed method is cubic with the number of transmit antennas. A flexible complexity#x002F;performance trade-off can be obtained by varying the subspace size. Even for small subspace dimensions, close to ML performance can be achieved, resulting in low complexity. The performance of this method is assessed in terms of bit error rate and achievable mutual information, and compared to conventional MIMO detection methods using the zero-forcing (ZF) or minimum mean squared error (MMSE) criterion, as well as the successive interference cancellation (SIC) approach. Furthermore, the convergence behavior of this iterative detection method is analyzed using a new variant of the EXIT-Chart based on the mutual information of equivalent hard output channels, i.e., a binary symmetric channel (BSC) model.