This paper maps the performance of various deep learning models, such as CNNs, RNNs, and hybrid models for real-time cyber threat detection. While Intrusion Detection System (IDS) has become integral to cybersecurity, several researches have been conducted with the aim of investigating different deep learning algorithms in order to enhance the accuracy of detection. It is an effort to review 48 existing papers that are focused on the effectiveness of deep learning models in the development of IDS to enhance its accuracy and efficiency. This gives a critical analysis of some of the normally used algorithms: deep learning and hybrid models. It identifies the most effective techniques that improve the models and suitable datasets to train them. The top-contributing algorithms, some of which are hybrid models including CNN and LSTM, models with a combination of machine learning and deep learning methods with accuracies of about 98% and 99% and their contributions toward a better detection performance when applied to specific datasets were pointed out. Although the review covers a wide range from algorithms to datasets, it also openly admits a number of issues: a potentially biased selection of studies, along with variability in dataset quality, may have an impact on the generalization of these results. Along with that the limitations of widely used datasets, like NSL-KDD, and topics like adversarial machine learning attacks have also been discussed. The research challenges for the future involve addressing model limitations and data scarcity issues, integrating advanced techniques, and exploring novel datasets to create more robust cybersecurity solutions. Also it provides insight into how IDS capabilities can be used through deep learning and the future scope for research in this field.