In this paper we present a comprehensive study on hate speech detection in Bengali (Bangla), a low-resource language with significant online presence. We explore the potential of large language models (LLMs) such as GPT-4, Qwen, and DeepSeek in identifying hate speech from social media content, including transliterated and code-mixed text. Using a consolidated dataset combining multiple public hate speech corpora, we evaluate LLM-based prompting and fine-tuning strategies alongside traditional deep learning and transformer models. Our findings show that fine-tuned LLMs like DeepSeek-67B and GPT-4 consistently outperform smaller models, achieving macro-F1 scores above 90%, with an ensemble of DeepSeek and XLM-R reaching 91.2%. These models also demonstrate stronger robustness to domain variation and better confidence calibration. This work highlights the value of LLMs in enhancing multilingual open-source threat intelligence and sets a new benchmark for hate speech detection in under-resourced language settings.