In this paper, we propose a novel architecture for a soft-output stochastic detector in multiple-input, multiple-output (MIMO) systems. The stochastic properties of this detector are studied and derived in this work, and several complexity reduction techniques are proposed to significantly reduce its cost from an architecture-implementation perspective. We also propose an efficient architecture to implement this detector. Finally, this detector is incorporated into an iterative detection-decoding structure, and through simulations, it is shown that the overall frame error rate (FER) performance and complexity is of the same order as that of the conventional K-best sphere detector.