This paper contributes to the field of inverter modelling for large-scale simulations by introducing a novel Model-Order Reduction (MOR) method based on singular perturbation. Motivated by the timescale separation between the fast and slow dynamics in an inverter-based power system, the proposed nonlinear MOR concept extends on the existing zero- and first-order reduction methods by combining the low computational burden of the former approach with the higher accuracy of the latter one. As a result, such hybrid MOR technique preserves the slow system dynamics of the full-order model, while simultaneously capturing the impact of the removed fast states on slow variables. Moreover, we introduce several improvements to the existing first-order MOR in order to make it tractable and more efficient when applied to a realistic full-order inverter model. The novel hybrid approach is applied to both grid-forming and grid-following inverter control schemes, and compared against existing reduction methods from the literature. The results showcase a better time-domain performance of the hybrid method during transients, while having a negligible increase in computational requirements compared to the traditional zero-order approach.