Stress, particularly mental stress, is a state of mind that results from one’s perception of the external or internal environment, which in turn leads to the psychological response of anxiety or distress. Often, social media is used as a platform to express this kind of stress in the form of extreme statements. It is crucial to identify and moderate such content on social media platforms, as the virtual space is accessed by millions of people from around the world. This paper develops and evaluates machine learning models to identify stressed and non-stressed comments from social media texts. We utilized various Large Language Models (LLMs) like Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Pretraining Approach (RoBERTa), Flan-T5 and Bidirectional Auto-Regressive Transformers (BART). We also experimented with traditional models to prove the superiority of LLMs over traditional models such as K-Nearest Neighbour, Logistic Regression, Decision Tree Classifier, Support Vector Machine and also ensemble models including Random Forest, Meta Stacking, Gradient Boosting and Bootstrap Aggregating. Additionally, we examined the impact of different vectorization techniques, including TF-IDF and Word2Vec on the traditional models. These models were applied and evaluated on a dataset consisting of Telugu language comments, offering important insights into the effectiveness of the approaches used.