[Purposes] To address the challenges of significant nonlinear characteristics in steel truss girder bridges under small-probability failure and the difficulty in solving implicit performance functions, a reliability analysis method for small-probability failure based on an improved Multi-gate Mixture-of-Experts (MMoE) surrogate model is proposed. [Methods] First, an improved MMoE surrogate model is constructed by optimizing the multi-task feature fusion mechanism. Then, using material parameters of the steel truss girder bridge as inputs and structural responses as outputs, the model is trained and validated, establishing structural performance functions that consider dual failure modes of cable stress and girder deflection. Finally, reliability analysis of a steel truss girder bridge is conducted by integrating the Monte Carlo Simulation (MCS) method with the improved MMoE surrogate model. [Findings] The results show that compared to BP neural networks and the original MMoE model, the improved MMoE model converges when the sample size exceeds 300, achieving a coefficient of determination R2 of 0.9872 on the validation set, with significantly superior prediction accuracy and stability. The errors of random test points on the local response surface follow a normal distribution centered at zero, confirming the model"s capability to accurately capture nonlinear structural responses. [Conclusions] The reliability indices of both cables and main girder components meet engineering safety requirements. Cables near the mid-span and main girder sections in the central span exhibit relatively lower reliability, whereas cables near the piers and main girders in side spans show higher reliability. Components with lower reliability should be prioritized in practical engineering safety assessments and maintenance decisions.