Large Language Models (LLMs) are predominantly trained on English data, leading to significant performance challenges for low-resource languages. This study focuses on Urdu as a case study to explore how LLMs process low-resource languages. We find that when prompted in a low-resource language, LLMs primarily reason internally in English, and this internal reasoning is more coherent than the generated text in the target language. This contrast highlights a gap between the model’s ability to comprehend low-resource languages and its ability to generate text in them. By analyzing these mechanisms, this work underscores the need for targeted improvements to enhance LLM performance for low-resource languages such as Urdu.