Low-light image enhancement remains a key challenge in vision domain, particularly when converting images from the RAW color space to the RGB domain. Recently, Transformers have become increasingly prevalent in vision tasks owing to their ability to capture extensive contextual relationships. However, their high computational complexity and vulnerability to noise still leave room for improvement in image enhancement. Therefore, we propose a hybrid U-Net model to improve low-light images using RAW inputs and a CNN-Mamba fusion. The model adopts a simplified three-layer U-Net, which takes a RAW image converted into a four-channel RGGB format via PixelUnshuffle as input. We then perform multi-scale feature extraction through a specially designed local-global module, MCMBlock, which integrates both CNN and Mamba mechanisms. A PixelShuffle operation is subsequently employed to reconstruct the RGB output, yielding enhanced image quality. With a significant reduction in parameter count and computational overhead, RCMUNet surpasses existing state-of-the-art methods, achieving 0.94 dB and 0.31 dB PSNR improvements on the MCR and the FiveK public datasets.