This paper presents the UPES_NLP submission to the WMT 2026 Low-Resource Arabic–Asian Machine Translation Shared Task. We used the NLLB-200 Mixture-of-Experts model for translation inference and GPT-5-mini with entity-aware prompting to preserve named entities. Our system participated in both Primary and Contrastive tracks across multiple language pairs. It ranked 3rd in Arabic→English with a BLEU score of 32.08 and 2nd in Arabic→Indonesian with a BLEU score of 22.77. The results show that the proposed approach performs competitively across low-resource Arabic–Asian translation tasks.