To address the quadratic computational cost of self-attention in Offline Handwritten Text Recognition (HTR), we propose HTR-HSS, a parameter-efficient hybrid architecture for efficient long-sequence modeling. The proposed model combines a lightweight CSP-UNet backbone for multi-scale visual feature extraction, a temporal downsampling module for sequence compression, and a hybrid sequence encoder composed of bidirectional Mamba blocks with sparsely inserted self-attention layers. This design enables efficient long-range dependency modeling while preserving local feature alignment. Without relying on pre-trained weights or external data, HTR-HSS achieves competitive recognition performance on the IAM, READ2016, and LAM benchmarks. In addition, the proposed model exhibits near-linear inference scaling with increasing input sequence length. Experimental results demonstrate that combining bidirectional Mamba with sparse self-attention provides an effective and practical solution for offline HTR, achieving a favorable balance among recognition accuracy, computational efficiency, and model size.