Deploying a Local Language Learning Assistant Using a Small Large Language Model
Konstantin A. Aksyonov, Lina Sun, Igor A. Kalinin, Olga P. Aksyonova, Elena K. Aksyonova · 2025
This paper investigates the feasibility of deploying a localized language learning assistant using small-scale large language models (LLMs). The proposed tool aims to enhance language learning efficiency by automating the organization and summarization of teaching notes. We leverage the lightweight and locally deployable nature of models such as Qwen, Ollama, and DeepSeek, combined with Retrieval-Augmented Generation (RAG) and Prompt technologies, to improve performance without additional fine-tuning. The study evaluates the models' inference speed and accuracy on consumer-grade hardware, focusing on tasks such as pronunciation error correction, advanced vocabulary extraction, and sentence structure improvement. In this work, five models were tested that could be deployed on consumer-grade computing devices. Experimental results demonstrate that the DeepSeek R1-8B model achieves high accuracy in note summarization, significantly reducing teachers' workload. However, limitations in semantic understanding and parameter size highlight areas for future optimization. This work provides a bit of data-referenced advice for the development of efficient, localised AI-driven language learning tools.