This study presents a comparative analysis of the zero-shot grading capabilities of three leading open-source large language models (LLMs). DeepSeek-7B, Gemma2-9B, and Llama3-8B. We evaluated the models on a dataset of 2,550 student responses from undergraduate business courses using a structured rubric-based prompting strategy. Our quantitative and statistical analyses revealed significant performance differences. Gemma2-9B demonstrated the highest alignment with human graders, achieving a Pearson correlation of 0.834, although it exhibited a slight positive bias. Conversely, DeepSeek-7B’s performance was substantially limited by output clamping, preventing it from scoring high-quality answers accurately. These findings highlight that while modern LLMs can be effective for automated assessment, they possess distinct biases and limitations that require careful model selection and domain-specific validation. This research contributes a robust methodology for zero-shot evaluation and provides critical insights for the practical implementation of AI in educational assessment.