In recent years, considerable efforts have been made to digitize historical manuscripts, providing scholars and end users with direct access to these images. Nonetheless, the processing of historical Arabic manuscripts continues to pose a significant challenge. This study aims to explore subword recognition in historical Arabic manuscripts through the use of end-to-end deep learning systems. The Convolutional Neural Network (CNN) is utilized in the phase of feature extraction. Bidirectional Long Short-Term Memory (BLSTM) and Bidirectional Gated Recurrent Unit (BGRU) models are employed utilizing a character model approach, with Connectionist Temporal Classification (CTC) serving as the decoding mechanism. By employing the IBN SINA database and integrating data augmentation, the BGRU attains the minimal Character Error Rate (CER) of 6.98% and the maximal accuracy rate of 91.52%. Our proposed method could accurately recognize subwords in historical Arabic manuscripts at both the character and subword levels without requiring segmentation, thereby overcoming several related challenges.