Hyperspectral images play a crucial role in precision agriculture, environmental monitoring or ecological analysis. However, due to sensor equipment and the imaging environment, the observed hyperspectral images are often inevitably corrupted by various noise. In this study, we proposed a truncated diffusion model, called TDiffDe, to recover the useful information in hyperspectral images gradually. Rather than starting from a pure noise, the input data contains image information in hyperspectral image denoising. Thus, we cut the trained diffusion model from small steps to avoid the destroy of valid information.