Although support vector machine (SVM) is an important algorithm, known hyperparameter tuning methods are typically time-consuming and susceptible to the influence of noise samples in large datasets. In particular, the selection of SVM hyperparameters is especially challenging in distributed learning. Therefore, this paper proposed a novel linear kernel SVM based on non-hyperparameter tuning. Its core idea was to train multiple SVM models using different regularization hyperparameters, and then integrate them into a final SVM model. To further increase the diversity of the resulting SVM models, Markov sampling was employed to generate different training subsets prior to training each SVM model. This paper derived the SVM based on non-hyperparameter tuning (SNHT) algorithm and proved its consistency. As an application, SNHT was applied to distributed learning. The performance of SNHT was validated through experiments on benchmark datasets.