ConMeZO is a novel zeroth-order optimization method designed to fine-tune large language models without relying on backpropagation, addressing the slow convergence challenges typical in high-dimensional parameter spaces. This adaptive descent-direction sampling technique improves the efficiency of finetuning billion-scale models, making it particularly relevant for developers working with expansive neural networks. By circumventing the need for gradient calculations, ConMeZO offers a promising solution for optimizing models where traditional gradient-based methods may be computationally prohibitive or infeasible. The approach has been proposed and positively received in recent research, highlighting its potential to enhance large-scale model adaptation. Its application scope includes scenarios where gradient-free optimization is advantageous, thus broadening the toolkit available for machine learning practitioners. Overall, ConMeZO represents an innovative step forward in the optimization of large language models, with benefits that could impact both research and practical deployment.