A new iterative technique for suppressing multiuser interference in uncoded DS-CDMA systems is proposed. In the new approach semi-hard decisions are taken at the output of the linear MMSE detector. This involves taking hard decisions only for symbols which satisfy a reliability criterion, while the rest are left unaltered in their soft form. The semi-hard estimate is subsequently used for parallel interference cancellation, which is then followed by additional linear MMSE filtering. This procedure is repeated a small number of times. Simulation results show that the proposed detector is capable of approaching closely the single-user performance in a fully loaded multiuser scenario and this is achieved within 5 to 10 iterations. In the same system scenario the new technique outperforms the probabilistic data association (PDA) algorithm, which is considered as a state-of-the art solution. ; A new iterative technique for suppressing multiuser interference in uncoded DS-CDMA systems is proposed. In the new approach semi-hard decisions are taken at the output of the linear MMSE detector. This involves taking hard decisions only for symbols which satisfy a reliability criterion, while the rest are left unaltered in their soft form. The semi-hard estimate is subsequently used for parallel interference cancellation, which is then followed by additional linear MMSE filtering. This procedure is repeated a small number of times. Simulation results show that the proposed detector is capable of approaching closely the single-user performance in a fully loaded multiuser scenario and this is achieved within 5 to 10 iterations. In the same system scenario the new technique outperforms the probabilistic data association (PDA) algorithm, which is considered as a state-of-the art solution. The computational complexity of the proposed detector is of the same order as of a standard linear detector.