arXiv:2611.24553v1 Announce Type: new Abstract: Using real-world massive MIMO measurements from the LuViRA dataset, we present a system-level solution combining Mamba-based model development and hardware acceleration for indoor localization. Three Mamba-based models are explored: the original Mamba architecture, CMamba, a channel-correlation-enhanced state-space model designed to exploit correlations across input channels, and our proposed Slim-Mamba. Hardware implementation and bottleneck analysis are then used to guide further model simplification and optimization. The resulting Slim-Mamba simplifies the original Mamba selective state-space block into a lightweight gated recurrent update, preserving temporal state modelling and output gating while reducing the complexity of the selective-scan computation. The deployed model contains 816k trainable parameters and has a model size of 3.40MB under FP32 precision while maintaining high localization accuracy. The corresponding accelerator provides a reconfigurable architecture with a highly parallel MAC array optimized for matrix-vector operations and fine-grained pipelining, improving resource utilization and reducing inference latency. The hardware-aware fixed-point model achieves a mean localization error of 0.14m, closely matching the floating-point baseline. Implemented on a Zynq UltraScale+ ZCU102 FPGA, the four-block backbone has an estimated serial latency of 212.04us per token at 100MHz.