Keystroke biometrics offers a promising alternative to traditional authentication methods by analyzing typing behavior for identity verification. This study introduces TypePlus, a novel deep learning architecture designed to improve free-text keystroke authentication without relying on transformer-based models. TypePlus integrates a weighted attention pooling mechanism to dynamically emphasize relevant keystroke features and an embedding layer for keycodes, enhancing generalization across diverse typing styles. Furthermore, we evaluate multiple loss functions, including Contrastive Loss, Euclidean and Manhattan Triplet Losses, and a novel Filtered Mean Triplet Loss. Experimental results on the Aalto University Keystroke Dataset demonstrate that TypePlus significantly outperforms previous models, achieving a 2.86% Equal Error Rate (EER). These findings underscore the potential of lightweight, non-transformer architecture in keystroke biometrics.