King and Minister Chess is a popular asymmetric board game in Tibet, Qinghai, Sichuan, and Gansu. We propose a Vision Transformer-based model to capture long-range dependencies and global context. The model utilizes a transformer encoder to process features such as White’s move, Black’s capture, and board state, thereby predicting value and policy. We also introduce a game evaluation algorithm that combines expert knowledge with Monte Carlo Tree Search to reduce computational overhead and enhance decision quality. A self-play randomization strategy is also proposed to balance training data, boosting white’s win rate. The model achieves a higher win rate against the reproduced state-of-the-art model. It also performs well against the intermediate human player.