In this work, we introduce a new model order reduction strategy, called Collocation-based Model Order Reduction (cMOR), proposed as an alternative to traditional projection-based approaches. Instead of computing the reduced solution through a projection of the full model, cMOR evaluates the solution only at a selected set of influential collocation points, identified using hyper-reduction techniques. The method preserves the usual two-stage workflow: an offline phase, where a reduced basis is built from snapshot data, and an online phase, where the high-fidelity model is sampled locally before reconstructing the full solution. We employ cMOR within the framework of nonlinear approximation manifolds, which extend classical reduced spaces by introducing nonlinear structures. This allows the method to better capture complex solution behaviours (e.g., multi-scale phenomena) while maintaining the simplicity and flexibility that make cMOR compatible with existing computational models.