Various models are available for large-scale evapotranspiration (ET) estimation. The performance of these models generally differs due to their differences in forcing data, model structure (mathematical representations of the ET process), and parameter estimation. Model comparison is the most straightforward way to identify the strengths, weaknesses, and uncertainty sources of a model. In this study, three widely used remote sensing ET models were considered: the air-relative-humidity-based two-source (ARTS) model, Priestley-Taylor Jet Propulsion Laboratory (PT-JPL) model, and Penman-Monteith-Leuning (PML) model. The three models were evaluated based on measurements from 12 eddy-covariance (EC) towers and water-balance-based ET estimates from 286 basins. The evaluation results indicate that the PML model performs the best on both the site and basin scales. The advantage of the PML model may be due to (i) the land-cover-based parameter configuration and (ii) the consideration of differences in the responses of soil evaporation and transpiration to soil water deficit. We also investigated the consistencies and differences between the models in simulating the spatiotemporal variation and component partitioning of ET, that is, interception loss (Ei), soil evaporation (Es), and transpiration (Et). The three models show high consistency in estimating nationwide multi-year average ET (416.3–438.2 mm/yr) and its spatial pattern but large discrepancies in ET trends (0.10–0.98 mm/yr2) and component partitioning. The PT-JPL model considerably overestimates the ratio of Ei/ET, thereby underestimating the ratio of Es/ET because of the negative correlation between Ei and Es. The ARTS model presents better applicability in grasslands than in croplands or forestlands. This may be because its parameter value (constant for all biomes) and the water constraint scheme are more suitable for grasslands. Finally, we proposed specific modifications to address the potential issues of each model.