The brain-computer interface (BCI) enables the translation of brain signals into commands for device control, playing a critical role in rehabilitation and communication technologies. Among various BCI applications, P300 spellers have gained significant attention for their reliability in character recognition. However, traditional BCI systems face challenges in cross-subject scenarios due to individual variability, which often degrades model performance and requires large training datasets for adaptation. To address the challenge of cross-subject variability in P300 spellers, we propose a novel multiscale convolutional network, MLNet, which incorporates a lightweight attention (LA) mechanism, coupled with a transfer learning-based few-shot fine-tuning (FT) strategy and prototype learning. Our approach effectively reduces the impact of individual variability, thereby enhancing performance cross-subjects. The LA module dynamically assesses the importance of neurons without introducing additional model parameters, thus improving the model’s adaptability to varying subject-specific characteristics. Moreover, few-shot learning enables the model to fine-tune with minimal data, while prototype learning further enhances generalization by leveraging representative examples from the dataset. Experimental results demonstrate the efficacy of our approach in improving cross-subject performance. After fine-tuning with only 10% of the available data, the average accuracy increased to 84.94%. These results suggest that our method significantly reduces the need for extensive calibration and accelerates the model adaptation process, offering a robust and practical solution for P300 BCI spellers.