Abstract: In recent years, large language models have shown remarkable capabilities in program generation and complex instruction understanding, enabling a new human-machine collaboration paradigm for CNC programming. This paper proposes a human-machine collaborative CNC programming method based on large language models. Part geometric features and process information are first structurally extracted and input into the model via unified natural language templates, allowing the model to generate machining programs. These programs are refined through human review and simulation verification to form a closed-loop programming workflow. Three collaborative programming templates are developed for contour machining, hole machining and thread machining. Taking a complex bearing seat as an example, this paper illustrates the full process of drawing information extraction, process planning, node coordinate calculation and input sentence construction, and systematically analyzes typical deviations and error-prone points in large-model-based human–machine collaborative programming. On this basis, it summarizes the methods and existing issues of human-machine collaborative CNC programming and puts forward directions for further improvement.