Large Language Models (LLMs) have demonstrated excellent performance across numerous AI fields. However, due to constraints related to latency and storage, LLMs are primarily being run on the cloud. Given the urgent needs of users to conserve bandwidth, reduce costs, and enhance privacy, researches on deploying LLMs on edge devices have gradually emerged as a hot topic. Yet, the limitations of edge devices in terms of cost-efficiency, restrained computing resources and multiple data types pose new challenges for evaluating LLMs on edge. This paper proposes a use case for evaluating large language models on edge devices based on Evaluatology method. By defining Evaluation Conditions (ECs), the conditions set for evaluation to ensure consistency for the same subject and comparability for different subjects, this paper clarifies the key metrics for evaluating LLMs on edge. We define comprehensive evaluation conditions, including model performance (such as inference speed, inference accuracy, memory usage, etc.) and edge device performance (such as device memory and cost-effectiveness). Simultaneously, we deploy currently popular large language models, including MiniCPM, Octopus, Phi-3, and others, on various edge devices. Utilizing these evaluation conditions, we conduct systematic evaluation experiments to compare and analyze the performance of different models across diverse edge devices. By using evaluatology, we can not only more accurately evaluate the performance of large language models on edge devices, but also provide help for optimizing their performance on edge.