Keystroke dynamics is a well-established behavioral biometric for continuous user identification. However, most existing models are device-specific, leading to performance degradation when users switch between desktop keyboards and touch-based devices. This presents a challenge in today’s multi-device environments, where seamless and secure identification is essential. In this work, we propose a unified Transformer-based model for cross-device user identification using keystroke dynamics. Leveraging transfer learning, our approach learns a shared representation of typing patterns from both desktop and mobile inputs, enabling robust identification across heterogeneous devices without requiring separate models. Experiments on the Keystroke 136M and Aalto Mobile datasets show that our model achieves Equal Error Rates (EERs) of 2.45 on desktop, 1.76 on mobile, and 2.63 in cross-device scenarios. These results confirm the effectiveness of our approach in achieving reliable, device-agnostic identification while improving usability and security in real-world, multi-platform settings.