Recently, hyperspectral image (HSI) classification with limited training samples has been well studied, where the metric-based few-shot classification or cross-domain few-shot classification methods have shown good performance. However, overcoming the overlapping distribution of classes and outliers solely through limited labeled samples remains a challenging problem. Moreover, too few labeled samples may lead to susceptibility to intra-class variations during cross-domain classification, and over-reliance on HSI modality constrains the accuracy of prototype representations to a certain extent. As such, a language dynamic-guided spatial-spectral network (LDSNet) framework is proposed for cross-domain few-shot HSI classification. Firstly, a multibranch spatial-spectral feature dynamic extraction model is designed to explore the local fine details, global structures, and spectral correlations, where the class text information is considered as a guidance to drive the whole model to extract the class-wise spatial-spectral features. Then, a cross-model alignment module is introduced to improve the class prototypes from spatial-spectral features through image-text matching, increasing the inter-class distances. Finally, in response to the intra-class variations and interference caused by limited labeled samples, a combination method of integrating prompt consistency learning and self-similarity masking learning is further developed. Hence, stable feature representations that do not vary with sample distribution can be generated, which ensures the representation consistency between different samples, improving the classification performance of target domain. Extensive experiments conducted on four HSI datasets demonstrate the superiority of our LDSNet framework in comparison with other state-of-the-art methods for cross-domain few-shot HSI classification.