Large language models (LLMs) have taken the world by storm, demonstrating impressive abilities in languages like English and Chinese. But what about less common languages like Mongolian? A new research paper introduces MM-Eval, a benchmark designed to test how well LLMs understand and reason in Modern Mongolian. Researchers wanted to understand not just *if* LLMs could process Mongolian, but *how* deeply they could grasp the language's nuances. They categorized LLM capabilities into two key areas: language abilities (syntax and semantics) and cognitive abilities (knowledge and reasoning), creating a hierarchical testing framework. Think of it like a language proficiency test for AI. The team used a Mongolian language textbook as the basis for their syntax and semantics tests, scrambling word order to check for grammatical understanding and testing vocabulary comprehension. For knowledge and reasoning, they drew from existing datasets like WebQSP and MGSM, adapting them to Mongolian to assess general knowledge transfer and problem-solving skills. The results? While LLMs like GPT-4 showed some proficiency in basic Mongolian grammar and transferred some general knowledge, they struggled with deeper semantic understanding and complex reasoning tasks. Interestingly, all models tested, even powerful ones, performed poorly on mathematical reasoning, suggesting a significant gap in their ability to apply logic in this context. The creation of MM-Eval marks a crucial step forward in understanding how LLMs handle less common languages. It highlights the challenges of transferring knowledge and reasoning abilities across different linguistic landscapes and opens the door for future research focused on improving LLM performance in low-resource languages like Mongolian. The findings also underscore the broader challenges of ensuring AI benefits all languages, not just the most widely spoken ones. As LLMs continue to evolve, benchmarks like MM-Eval will be essential for measuring their true multilingual capabilities and guiding the development of more inclusive and truly global AI systems.