<p>The identification of hate and offensive content in multilingual social media posts remains challenging due to the high cost of manual annotation and the complexity of code-mixed language. Most existing methods depend on large amounts of fully labeled data, which is not practical for real-world moderation systems. In this work, we propose a BERT-based active learning framework that reduces the annotation requirement to only 25% of the training data, while maintaining or improving detection performance across English, Hindi, and German. Unlike prior work that primarily focused on monolingual English, our approach extends active learning to multilingual and code-mixed settings using a pool-based uncertainty sampling strategy over iterative queries. Experimental evaluation on HASOC-2019 and HASOC-2020 benchmarks shows that the proposed framework achieves substantial performance improvements. For example, on English Task 1 of HASOC-2020, our model reaches a weighted F1-score of 0.90, outperforming the best-reported baseline (0.51) by 39 percentage points, while using only one-fourth of the labeled samples. Similarly, for German Task 1, our model achieves 0.81 weighted F1-score (macro F1 = 0.76) compared to the earlier best of 0.52. These results demonstrate the efficiency and effectiveness of active learning for multilingual hate speech detection. The key novelty of this work lies in its integration of transformer-based active learning with multilingual datasets, offering both data efficiency and state-of-the-art performance with reduced labeling costs.</p>