Key points are not available for this paper at this time.
The rapid integration of large language models (LLMs) into education has shifted research focus from questions of capability, such as what LLMs can do and how accurately—to questions of deployability, including how they can be operated effectively for many learners at once. In school environments, system reliability, scalability, and real-time responsiveness are critical, as delays or interruptions can directly reduce learner engagement, particularly during synchronous activities. This study evaluates the performance of open-source LLaMA models ranging from 1 billion to 70 billion parameters across one-, dual-, triple-, and quad-GPU configurations suitable for educational settings. Performance is assessed using four key indicators: success rate (percentage of completed requests), generation speed (tokens per second), throughput (completed responses per second), and latency (time until full response generation). These metrics were measured under progressively increasing numbers of simultaneous users to identify system capacity limits and trade-offs between model size, responsiveness, and scalability. The results indicate that smaller models (1B–3B) deliver faster, more stable performance under concurrent use, while larger models (8B–70B) experience substantial slowdowns and reduced reliability, even on high-end GPU systems. These findings suggest that effective educational deployment should prioritize empirical performance and infrastructure compatibility over model size alone. The paper concludes by proposing a practical framework to guide educators, administrators, and developers in selecting and configuring locally hosted GPU systems that balance model capability, response speed, and resource efficiency for real-time applications such as AI tutors, classroom chatbots, and automated feedback tools.
Chan et al. (Mon,) studied this question.