This study evaluates and compares the performance of two transformer-based multilingual language models: XLM-RoBERTa and MuRIL, for extractive question answering (QA) tasks in Hindi and English. The study focuses on the effectiveness of leveraging pre-trained transformer architectures through task-specific fine-tuning, with particular emphasis on accuracy, language performance, and computational efficiency. The datasets, collected and prepared for bilingual QA experimentation and made available through Kaggle, were divided into training and testing subsets using an 80:20 split. Model performance was evaluated using multiple complementary metrics, including Exact Match (EM), F1 Score, BLEU Score, ROUGE-1, ROUGE-2, and ROUGE-L, along with average response time per question. The experimental results demonstrate that fine-tuning can substantially improve QA performance when task-specific adaptation is beneficial, particularly for MuRIL and Hindi QA, whereas the English XLM-RoBERTa results show that direct evaluation of the pre-trained model can outperform the fine-tuned configurations. However, the experiments also reveal the risk of overfitting when excessive training epochs are employed, as reflected by increasing training loss and deteriorating validation performance. Quantization was additionally investigated as an optimization technique and demonstrated meaningful reductions in inference response time, although its effect on predictive accuracy was model and language dependent. Overall, the findings highlight the potential of multilingual transformer models for developing effective bilingual QA systems and provide practical insights into balancing model accuracy, generalization, inference efficiency, and computational requirements, particularly for Hindi and other relatively low-resource languages.