Facial Expression Recognition (FER) technology, as an external indicator of human emotions, holds significant practical value in real-world applications. However, issues such as lighting variations and facial occlusion in complex environments challenge model robustness. Additionally, existing deep learning methods often suffer from a large number of parameters and high computational complexity, making them difficult to deploy on edge devices. This paper proposes LLMamba-Net, a lightweight network featuring a backbone customized for FER tasks. The feature expression capability is enhanced through the design of Morphological Convolutional Channel Local Attention (MCCLA). Efficient co-modeling of local and global features is accomplished by designing the Linear Context Broadcast Mamba (LCB-Mamba). With just 5.11M parameters, LLMamba-Net achieves competitive results on several widely used datasets, offering an innovative solution for deploying high-accuracy expression recognition models on resource-constrained devices.