Sarcasm detection is an imperative undertaking within the realm of natural language processing, albeit one that poses considerable challenges when confronted with mash-up languages, characterized by the amalgamation of multiple distinct languages. In response to the intricacies of sarcasm detection in mash-up languages, with a specific focus on the Indonesian-English language mash-up, this study introduces the Hybrid Pretrained Word Embedding approach as a means to enhance sarcasm detection. The primary objective of this research is to augment the precision of sarcasm detection in mash-up languages by amalgamating suitable word embeddings tailored to the employed terms. The present study combines two prevalent pretrained word embeddings, i.e Glove and Fasttext, wherein Glove is utilized to extract semantic context vectors for English words, while Fasttext is employed to extract semantic context vectors for Indonesian words. The classification process in this research leverages the deep learning methodology known as Bidirectional Gated Recurrent Unit (BiGRU). To assess the efficacy of the proposed approach, an extensive dataset comprising sarcastic and non-sarcastic tweets, written in a hybrid language of Indonesian and English, is acquired from the Twitter platform. The results unequivocally demonstrate that the Hybrid Pretrained Word Embedding approach significantly enhances sarcasm detection in mash-up languages, attaining a commendable classification accuracy of 93.57% and an F-measure of 97.94%. By offering an effective methodology to identify sarcasm in mash-up languages, this study yields a substantive contribution to the field of natural language processing.