Abstract: The design of composites structures mainly relies on the safety factor method. However, due to the significant variability in the mechanical properties of composites, this approach can lead to overdesign or reliability issues, limiting their lightweight potential. Probabilistic methods have been introduced to address this, but reducing their computational cost requires numerical simulations that account for the randomness in composite mechanical properties. This study uses unidirectional carbon fiber (TZ700S-12K) fabric and epoxy resin to prepare laminated composites. By combining experiments and simulations, this research develops a numerical model to accurately predict the statistical distribution of the flexural properties of composites. Tests have been conducted to obtain discrete data on the strength and modulus of unidirectional composites, as well as interlaminar parameters, with statistical characterization performed using the Weibull distribution. Based on this foundation, two discrete finite element models for bending were established at the local and global scales. Within these models, material properties were spatially randomized through a self-developed Python script in ABAQUS, while intralaminar damage and interlaminar delamination were simulated by integrating the 3D Hashin failure criterion with a cohesive zone model. The local discrete bending model effectively captures the scattering characteristics of bending performance. Its predictions for bending strength and failure strain show strong agreement with experimental values, while the failure process exhibits progressive evolution featuring localized initiation and gradual expansion of damage. The material property assignment in the global discrete bending model does not effectively represent the local performance variations within and between layers, leading to overestimated predictions and an overall abrupt failure pattern during the damage process. The random multi-scale analysis framework developed in this study provides support for characterizing the variability and reliability design of bending properties in composites.