This paper introduces a novel hybrid convolution and attention network (HCANet) for hyperspectral image (HSI) denoising. The proposed model combines the strengths of convolutional neural networks (CNNs) and Transformers to enhance the denoising performance by simultaneously modeling both global and local features. The key contributions of HCANet include: 1. **Convolution and Attention Fusion Module (CAFM)**: This module captures both long-range dependencies and neighborhood spectral correlations, leveraging the strengths of convolutional and attention mechanisms. 2. **Multi-Scale Feed-Forward Network (MSFN)**: This network improves multi-scale information aggregation by extracting features at different scales, enhancing the denoising performance. The HCANet is designed as a U-shaped network with multiple CAMixing blocks, each containing a CAFM and an MSFN. The model uses 3D convolutions for low-level feature extraction and 2D convolutions for channel adjustment to manage complexity. The denoising process involves a noise residual map, which is subtracted from the noisy HSI to obtain the clean HSI. The model is trained using an $L_{1}$ loss function and a global gradient regularizer to improve denoising quality and reduce redundancy. Experimental results on two benchmark datasets (ICVL and Pavia) demonstrate that HCANet outperforms state-of-the-art methods in terms of denoising performance, as measured by PSNR, SSIM, and SAM metrics. The model effectively removes various types of complex noise, including Gaussian, impulse, and deadline noise, while preserving local details and maintaining moderate computational complexity.This paper introduces a novel hybrid convolution and attention network (HCANet) for hyperspectral image (HSI) denoising. The proposed model combines the strengths of convolutional neural networks (CNNs) and Transformers to enhance the denoising performance by simultaneously modeling both global and local features. The key contributions of HCANet include: 1. **Convolution and Attention Fusion Module (CAFM)**: This module captures both long-range dependencies and neighborhood spectral correlations, leveraging the strengths of convolutional and attention mechanisms. 2. **Multi-Scale Feed-Forward Network (MSFN)**: This network improves multi-scale information aggregation by extracting features at different scales, enhancing the denoising performance. The HCANet is designed as a U-shaped network with multiple CAMixing blocks, each containing a CAFM and an MSFN. The model uses 3D convolutions for low-level feature extraction and 2D convolutions for channel adjustment to manage complexity. The denoising process involves a noise residual map, which is subtracted from the noisy HSI to obtain the clean HSI. The model is trained using an $L_{1}$ loss function and a global gradient regularizer to improve denoising quality and reduce redundancy. Experimental results on two benchmark datasets (ICVL and Pavia) demonstrate that HCANet outperforms state-of-the-art methods in terms of denoising performance, as measured by PSNR, SSIM, and SAM metrics. The model effectively removes various types of complex noise, including Gaussian, impulse, and deadline noise, while preserving local details and maintaining moderate computational complexity.