Convolutional neural networks (CNNs) have been extensively applied to hyperspectral image (HSI) classification, known for their ability to capture local spectral-spatial features. However, CNNs face limitations in modeling long-range dependencies and global contextual information, which are crucial for the complex nature of HSI data. To address this shortfall, we propose a hybrid architecture that leverages the strengths of both CNNs and transformers. While CNNs excel in extracting local features, transformers provide the ability to model global interactions, thereby delivering a more holistic representation of the data. Despite their effectiveness, conventional transformers suffer from quadratic computational complexity due to the self-attention mechanism, making them resource-intensive. To mitigate this issue, we introduce a Gaussian-Kaiming Focused Linear Attention (GKFLA) mechanism in the transformer component of our model. In this design, the query (Q), key (K), and value (V) matrices are initialized using Kaiming initialization, enhancing the model’s capacity to extract meaningful features. Furthermore, a Gaussian matrix is added to the query to reduce overfitting and introduce greater non-linearity, optimizing the soft-max operation. This focused linear attention mechanism not only reduces computational overhead but also accelerates the model, making it more efficient than traditional soft-max-based transformers without sacrificing accuracy. We evaluate the performance of this approach on three benchmark HSI datasets—Indiana Pines (IP), University of Pavia (PU), and Salinas (SA)—demonstrating superior classification results with improved computational efficiency by combining CNN and transformer capabilities.