Abstract: To address the problems of low generalization and high testing costs of most current residual strength models for composites, a probabilistic residual strength model that accounts for the effect of stress level and is independent of stress level was proposed. Firstly, the normalized strength reserve was defined and a deterministic residual strength model was derived based on the normalized strength reserve. Then, a fatigue life probability model was coupled into the deterministic residual strength model, and then a new residual strength probability model was derived. Finally, the accuracy and applicability of the proposed probabilistic residual strength model was verified using constant-amplitude and variable-amplitude residual strength experimental data from the open literatures. The results show that almost all the constant amplitude experimental data points are distributed between the upper 95% confidence limit and lower 5% confidence limit of the prediction curves, and the prediction curves with 50% reliability have high goodness-of-fit values for the experimental data: 0.94, 0.84 and 0.97. The proposed model accurately describes strength degradation at multiple stress levels using only one set of model parameters, with sufficient consideration of the statistical characteristics of the residual strength of the composite. The relative error between the predicted values of the proposed model and experimental values for both ascending and descending variable-amplitude loading is less than 6%.