<p>Salient object detection aims to identify visually prominent regions in RGB-D images. Recent state-space models have shown strong potential for this task, but many high-capacity or Mamba-based RGB-D methods still underexplore modality-specific capacity allocation. RGB images usually contain dense semantic and texture information, whereas depth maps provide sparse geometric structure and may include measurement noise. To improve the accuracy-efficiency trade-off, we propose HMamba, a heterogeneous gated state-space fusion network for RGB-D salient object detection. HMamba assigns a high-capacity VMamba encoder to RGB input and a lightweight EfficientVMamba encoder to depth input. This design matches model capacity to modality information density. We further introduce a gated bidirectional state-space fusion module and a hierarchical feature aggregation decoder to integrate cross-modal cues and multi-scale features. Experiments on eight RGB-D SOD benchmarks show that HMamba achieves competitive detection accuracy with favorable computational efficiency. Compared with the contemporary Mamba-based baseline, HMamba reduces model size by 51.2% and improves inference speed by 24.8%, while maintaining or improving saliency detection performance across most benchmarks.</p>