Enzymes， as efficient and environmentally friendly biocatalysts， are widely applied in medicine， food， chemical， agricultural， and other industries. However， the wild-type enzymes often suffer from poor thermostability， narrow substrate spectrum， and limited activity. Protein language models (PLMs)， inspired by natural language processing， are trained on large-scale protein sequence and structure datasets through self-supervised learning to capture evolutionary patterns linking sequence， structure， and function， thereby showing great potential in enzyme engineering. This review systematically summarized representative PLM architectures and training strategies， and highlighted their recent applications in zero-shot and few-shot prediction， enzyme function prediction， and de novo design. Specifically， PLMs have been used for mutation effect prediction and catalytic performance optimization， and to accelerate iterative evolution when integrated with automated platforms， thereby enhancing enzyme thermostability， activity， and substrate adaptability. Moreover， multimodal representations and few-shot learning approaches have improved task-specific prediction accuracy， while PLMs have also enabled the design of novel functional enzymes. Finally， the challenges of model scalability， generalization， and integration with biophysical knowledge were discussed， along with future prospects of PLMs for controllable functional protein design and industrial applications.