Abstract: Experimental study is carried out to study the reasonable statistical model for the random fracture toughness data of LZ50 axle steel. It is found that the material is brittle and a strict control of the manufacturing surface quality of railway vehicle axles is required. Meanwhile, it is shown that the existent normal distribution-based model is not useful to descibe the test data. The goodness-of-fit, consistence with failure physics, and safety of predictions in tail region are examined for six statistical distributions, i.e., three-parameter Weibull, two-parameter Weibull, normal, lognormal, extreme minimum value, and extreme maximum value. Results show that the extreme minimum value distribution is a reasonable model. A probabilistic method is constructed on the basis of the model. The probabilistic values for the toughness are given under the typical survival probabilities and confidences.