Text summarization is a Natural Language Processing (NLP) task that condenses textual content while preserving essential information. With the exponential growth in digital news content, efficient summarization techniques have become imperative. Transformers have revolutionized tasks in natural language processing, including text summarization, due to their self-attention mechanism and adeptness at capturing long-range dependencies. However, these transformers are not task-specific optimized, are sensitive to hyperparameter configuration, and are associated with significant training and inference costs, which pose a challenge. Therefore, the main objective of this study was to develop an improved transformer-based model optimized for text summarization of news articles to solve the stated limitations. A true experiment was adopted for the research design, specifically, the pretest-posttest control group design. Four transformers, Generative Pretrained Transformers (GPT), Text-to-Text Transfer Transformer (T5), Bidirectional and Auto-Regressive Transformer (BART), and Pre-training with Extracted Gap-sentences for Abstractive Summarization (PEGASUS), were subjected to a comparative stud on the CNN/DailyMail dataset. The T5 was identified as the baseline model due to its text-to-text framework and the encoder-decoder architecture. Subsequently, the T5 was improved through adaptive model quantization to reduce training and inference costs, hyperparameter optimization using manual tuning and gradient descent optimization to optimize its performance, and output layer modification by incorporating temperature and top-k sampling parameters to improve the quality of the generated summary. The improved T5 model was trained on the CNN/DailyMail dataset and evaluated on the XSUM and CNN/DailyMail datasets using ROUGE and BLEU metrics. The improved T5 model outperformed the baseline T5 across the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) and Bilingual Evaluation Understudy (BLEU) evaluation metrics on both benchmark datasets. The improved T5 model cumulatively outperformed the baseline T5 on CNN/DailyMail by 3.05% on ROUGE metrics and 4.2% on the BLEU score, and on the XSUM dataset by 7.27% in ROUGE metrics and 6.53% in BLEU score. The findings demonstrated that the improved T5 model outperformed the baseline across two evaluation metrics on both datasets, indicating improved summarization and generalization capabilities. This study has underscored the training and inference efficiency gains from quantization, the pivotal role of hyperparameter optimization in model convergence and generalization, and the effectiveness of output layer modification in improving text summarization using transformers. The improved T5 model can be deployed in real-world settings or applied in other domains beyond text summarization and utilized for multi-document summarization.