1. We studied the latencies and amplitudes of saccades to moving targets in normal human subjects. Targets underwent ramp or step-ramp motions. The goal was to determine how the saccadic system uses information about target velocity. 2. For simple ramp motion saccadic latency decreased as target speed increased. A threshold distance model, which assumes that the target has to move a minimum distance before saccadic processing starts, provided a good fit to the responses of all four subjects and explains discrepancies between previously published findings. 3. A double step experiment showed that target position may have some effect on saccadic amplitude when sampled approximately 70 ms before saccade onset, but it must be sampled at least 140 ms before onset for an accurate saccade to occur. 4. Saccades to simple ramp targets approximated the target position 55 ms before saccade onset. Based on our double step results, this is more compensation than possible by a simple position estimate and implies extrapolation of target motion by the saccadic system. The lack of complete compensation may be due to an underestimate of the target speed and/or of the saccadic latency. 5. A delayed-saccade paradigm resulted in saccades with a longer, constant latency and allowed longer viewing of target motion. These saccades accounted for all but approximately 20 ms of target motion, suggesting that with more processing time of target motion a better extrapolation may be generated. 6. In a step-ramp paradigm the target stepped in one direction, then moved smoothly in the opposite direction. Saccades in this paradigm could be made in either the direction of the step or in the direction of target motion: the direction and latency were determined solely by the time at which the target crossed the fixation point. This time must be calculated from target speed and position, implying that the saccadic system must use speed information to adjust latency or to cancel unnecessary saccades.