Abstract. Baseflow is an important component in hydrological modeling. Complex streamflow recession process complicates the baseflow simulation. In order to simulate the snow and/or glacier melt dominated streamflow receding quickly during high-flow period but very slowly during the low-flow period in rivers in arid and cold Northwest China, the current one-reservoir baseflow approach in SWAT (Soil Water Assessment Tool) model was extended by adding a slow reacting reservoir and applied to the Manas River basin in Tianshan Mountains. Meanwhile, a digital filter program was employed to separate baseflow from streamflow records for comparisons. Results indicated that the two-reservoir method yielded much better results than the one-reservoir one in reproducing streamflow processes, and the low-flow estimation was improved markedly. Nash-Sutcliff efficiency values at the calibration and validation stages are 0.68 and 0.62 for the one-reservoir case, and 0.76 and 0.69 for the two-reservoir case, respectively. The filter-based method estimated the baseflow index as 0.60, while the model-based as o.45. The filter-based baseflow responds almost immediately to surface runoff occurrence at onset of rising limb, while the model-based with a delay. In consideration of watershed surface storage retention and soil freezing/thawing effects on infiltration and recharge during initial snowmelt season, a delay response is considered to be more reasonable. However, a more detailed description of freezing/thawing processes should be included in soil modules so as to determine recharge to aquifer during these processes, and thus an accurate onset point of rising limb of the simulated baseflow.