Recent studies revealed that traditional techniques of understanding code changes are less effective than techniques directly prompting large language models (LLMs). However, current techniques utilizing LLMs heavily rely on commercial, large-scale ones such as GPT-3.5 and GPT-4, preventing their widespread practical deployment. This paper seeks to investigate the feasibility of deploying small-scale LLMs while maintaining comparable or superior performance to commercial and larger-scale LLMs in terms of code change understanding. To achieve this, we have developed a small yet high-quality dataset called HQCM which was reviewed, revised, and validated by five human experts. After finetuning small-scale (7B and 220M) LLMs via it, our evaluation has confirmed the significant profits brought by HQCM and has indicated that small-scale LLMs, after fine-tuning by HQCM, can achieve superior performance in change understanding for change summarization, change classification, and code refinement, compared to state-of-the-art baselines and larger-scale (>=70B) LLMs. This study supports the use of small-scale LLMs in industry or resource-constrained settings like embedded systems, distinguishing our work from others.