Malware is widely regarded as one of the most severe security threats to modern technologies. Detecting malware in the Internet of Things (IoT) infrastructures is a critical and complicated task. The complexity of this task increases with the recent growth of malware variants targeting different IoT CPU architectures since the new malware variants often use anti-forensic techniques to avoid detection and investigation. There-fore, we cannot utilize the traditional machine learning (ML) techniques that require domain knowledge and sophisticated feature engineering in detecting the unseen mal ware variants. Re-cent deep learning approaches have performed well on mal ware analysis and detection while using minimum feature engineering requirements. In this paper, we propose BERTDeep- Ware, a real-time cross-architecture malware detection solution tailored for IoT systems. BERTDeep- Ware analyzes the executable file's operation codes (OpCodes) sequence representations using Bidi-rectional Encoder Representations from Transformers (BERT) Embedding, the state-of-the-art natural language processing (NLP) approach. The extracted sentence embedding from BERT is fed into a customized hybrid multi-head CNN-BiLSTM-LocAtt model. This deep learning (DL) model combines the convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and the local attention mechanisms (locAtt) to capture contextual features and long-term dependencies between OpCode sequences. We train and evaluate BERTDeep- Ware using the datasets created for three different CPU architectures. The performance evaluation results confirm that the proposed multi-head CNN-BiLSTM-LocAtt model produces more accurate classification results with higher detection rates and lower false positives than a number of baseline ML and DL models.