Defects, especially for knots, seriously affect the bending strength of wood. Full scale specimens and clear specimens for bending test, cut from dimension lumbers of Larix gmelinii with different kinds of knots, were applied to evaluate the predictive accuracy of the bending strength and to improve the accuracy of the bending strength by multivariate regression analysis with density and knot information. According to the regression analysis results, the knot Ik/Ig of the weakest location for tension side and the average density ρa show the best predictive accuracy of bending strength for full scale specimens（ff）, with 0.753 of R2 and 4.71 of RMSE. Based on the ratio of regression parameter B/A of independent variables for density and knot information, the effect of knot and density are equally important in influencing ff. Finally, a simplified theoretical model that can be used to calculate bending strength for dimension lumber is constructed.