Zero-Shot Classification (ZSC) has indeed become one of the main topics of research in Natural Language Processing (NLP), as it allows the model to predict the test data belonging to classes that the model has never seen before with no access to any samples from such classes. In this paper, the capability of DeBERTa V3, one of the advanced pre-trained language models, is explored when it comes to zero-shot classification by means of Natural Language Inference (NLI). To this end, we concentrate on the application of MultiNLI dataset for demonstrating the ability of DeBERTa V3 in classifying the text into new classes. Semantic relations and contextualized embeddings allow DeBERTa V3 to perform well in tasks requiring generalization to unseen situations; It can classify cases based on premise-hypothesis pairs without requiring tuning to the specific task. To address these issues, we present a suggested procedure for preprocessing, feature engineering, andemos prompt tuning. The experimental setting measures the model’s usability via accuracy and top-k accuracy parameters. The evaluation also confirms that DeBERTa V3, with the suitable hypothesis template and prompt-tuning, surpasses typical zero-shot strategies, which significantly improve the reliability of practical applications in various NL abilities. This work demonstrates the potential of the zero-shot learning paradigm for practical scenarios where access to labeled data is limited, especially in the settings with multiple languages and domains.