Three models for torpor initiation were tested in rufous hummingbirds (Selasphorus rufus) during moult, when these birds appear to avoid the use of torpor. In model 1, the level of energy reserves at which torpor is initiated (the "threshold") remains constant throughout the night. In model 2, the threshold declines throughout the night, at a constant rate equivalent to the rate at which energy reserves are depleted during torpor. In model 3, the threshold declines at a rate equivalent to the rate of energy reserve depletion during torpor for most of the night, but at a higher rate (corresponding to the rate of energy expenditure during normothermia) during the final 2 h of the night, when these birds are usually normothermic. Model 1 predicts the most frequent and longest bouts of torpor, whereas model 3 predicts the fewest and shortest bouts. To determine the thresholds for each of 12 birds, food supply was manipulated to induce entry into torpor at different times on successive nights. Threshold slopes matched the predictions of model 3 most closely. Calculations comparing observed incidence of torpor with the predictions of model 1 show that the actual, time-dependent threshold for torpor initiation resulted in a 72% reduction in the number of torpor bouts compared with the number of torpor bouts that should have been initiated by a constant threshold. The advantage of a time-dependent threshold is that, although torpor is initiated when needed to prevent energy reserves from falling below a critical level, the amount of time spent in torpor can be minimized.(ABSTRACT TRUNCATED AT 250 WORDS)