Programming time is reduced in a non-volatile memory in a multi-pass programming process. In a first programming pass, high state cells are programmed by a sequence of program pulses to identify fast and slow high state cells, while lower state cells are locked out from programming. Once identified, the fast high state cells are temporarily locked out from programming while the slow high state cells continue being programmed to their final intended state. Further, the program pulses are sharply stepped up to program the slow high state cells. In a second programming pass, the fast high state cells are programmed along with the other, lower state cells, until they all reach their respective intended states. A time savings is realized compared to approaches in which all high state cells are programmed in the first programming pass.