Nano large language model of 125 million-parameter for STEM education

Large language models (LLMs) have achieved remarkable success on various natural language tasks, but their immense size often makes deployment on resource-constrained devices impractical. This paper presents Lightweight Nano LLM, a compact 125-million-parameter Transformer-based language model tailored for domain-specific question answering. Built on a GPT-2 small architecture, the proposed model is trained via knowledge distillation on a custom B.Tech CSE curriculum corpus ( $$\sim $$ 3 million tokens) to inject deep-domain knowledge. The study emphasizes techniques from prior literature, such as model compression and curated training data that enable small models to punch above their weight in coherence and accuracy. In evaluations, Lightweight Nano LLM demonstrates near-perfect accuracy on in-domain queries, outperforming other models of similar scale (e.g. GPT-2 small, DistilGPT-2, BERT-base) in this specialized task. However, the proposed model has outperformed all the above compared models with $$72.8\%$$ accuracy. The model’s compact size and focused training also allow it to run efficiently on consumer hardware (e.g. Apple M1), highlighting the promise of small LLMs for personalized and offline applications. The proposed work presents a detailed literature review, comparative analysis with related models, and an architectural diagram. The results show that with appropriate training data and design, smaller truly can be smarter in specialized settings for STEM disciplines.