Music source separation (MSS) for real-time applications faces challenges due to significant performance loss. In this paper, we propose Band-SCNet, a lightweight real-time model that bridges the performance gap between real-time and non-real-time models. Band-SCNet combines Sparse Compression with Cross-band and Narrow-band Blocks to reduce model size and complexity while improving performance. We also introduce the Compressed Self-Attention (CSA) Fusion Module, which enhances efficiency by reducing parameters. Experiments on the MUSDB18-HQ dataset show that Band-SCNet outperforms existing real-time models with an SDR of 7.79 dB, 2.59 million parameters by relaxing to a 92 ms latency, confirming its suitability for real-time MSS. The model achieves a balance between performance, latency, and model size, offering a promising solution for real-time music source separation.