Large language models (LLMs) have taken the AI world by storm, demonstrating impressive abilities in various tasks, including translation. But how well do they *really* perform when faced with the challenges of low-resource languages and limited computational power? New research puts LLMs to the test, specifically examining English-Thai translation, and the results are surprising. The study pitted a powerful LLM, Llama-3, against specialized translation models like NLLB. Under typical conditions, Llama-3 held its own, but the playing field changed dramatically when computational resources were restricted. As the researchers applied 4-bit quantization, simulating a low-power environment, Llama-3's translation quality plummeted. In contrast, the leaner, specialized NLLB models maintained their performance, outshining Llama-3 on standard metrics like BLEU3 and METEOR. This raises an important question: while LLMs boast impressive general abilities, are they truly the best tool for every job? This research suggests specialized models might be more robust and efficient, especially in resource-constrained scenarios, highlighting a potential hidden cost of relying solely on the largest LLMs. Furthermore, the study examined a unique challenge: code-switching, where a sentence mixes languages (like retaining English medical terms within a Thai sentence). Interestingly, Llama-3 initially showed a surprising strength in handling these mixed-language sentences, particularly in capturing the overall meaning, even if the precise wording suffered. However, this advantage also eroded under quantization. This research illuminates the complex trade-offs between general-purpose LLMs and specialized models. While LLMs offer impressive flexibility, their performance can be fragile when resources are limited. The future of AI likely lies in a balanced approach, leveraging the strengths of both generalist and specialist models to achieve robust and efficient language processing across diverse contexts.