Sumatriptan is indicated for the treatment of migraine attack and cluster headache; it is currently marketed as a subcutaneous injection, nasal spray, and oral tablet. New formulations are under consideration. The knowledge of sumatriptan absorption, combined with PK/PD information would help the design of more efficient formulations. In this perspective, we attempted to model the absorption of sumatriptan by population PK analysis. Data following administration by the intravenous (iv), the subcutaneous (sc), and the oral (po) route in healthy subjects were analyzed. A large database with full kinetic profiles was constituted. Sumatriptan was administered to 215 healthy subjects (iv, sc, and po) and to 143 migraine sufferers (po). The mean age was 31 years (18–86 years) in healthy subject population and was 38 years (18–65 years) in migraine patients. The mean weights were 74kg (54–104 kg) and 66 kg (38–136 kg) in healthy subjects and migraine patients, respectively, and the mean heights were 176cm (157–193 cm) and 164cm (152–183 cm) in healthy subjects and migraine patients, respectively. A NONMEM analysis was performed using a two-compartment disposition model. Oral absorption was modeled with a first-order input followed by a zero-order input. Less biased results were obtained using the FOCE method. The total clearance and the distribution volume at steady state were 71.2 L/hr and 94.5 L after iv dosing and 68.7 L/hr and 109 L after inclusion of the sc and po data. The absorption phase appeared to last for about 5 hr. The interindividual variability of the main PK parameters was low: It was around 20% for the total clearance and around 30% for the distribution volume at steady state. Although significant, the combination of age and height on clearance did not decrease considerably the interindividual variability of this parameter (decrease of 2.2%); nor was it possible to establish clearly if a migraine attack has an effect on drug absorption because of the sampling scheme during absorption. Simulations have shown that it would have been possible to estimate all the PK parameters with a data set reduced to one quarter of its actual number of samples.