This study addresses the limitations of twin support vector machines (TWSVM)—namely, insufficient flexibility and limited generalization capability when handling complex data structures—by proposing an improved TWSVM algorithm based on mixed kernel functions. By integrating local kernels (Gaussian and Laplacian kernels) with a global kernel (Sigmoid kernel), hybrid Gaussian and hybrid Laplacian kernels are constructed to enhance the model’s adaptability to multi-scale data features. Experimental results demonstrate that the proposed mixed-kernel approach outperforms single-kernel models across multiple datasets, significantly improving classification accuracy. The study also finds that the weight coefficients have a considerable impact on the performance of the hybrid kernels, and that reasonable configuration can effectively balance generalization and fitting capabilities. This research provides an effective kernel function construction strategy to enhance the performance of TWSVM in complex classification tasks.