This work investigates how small language models, especially those between 1B and 3B parameters, can be fine-tuned to handle programming tasks more effectively. With increasing interest in running Artificial Intelligence models on limited hardware, we tried out fine-tuning methods like Low-Rank Adaptation, Quantised Low-Rank Adaptation, and Unsloth to improve performance without needing expensive resources. Instead of building a new dataset, we used existing coding problem datasets from platforms like Leetcode, Codeforces etc. These datasets include challenges, test cases, and solutions, and they were useful for evaluating code generation and reasoning abilities. During fine-tuning, we faced common issues such as memory limits, long training times, and occasional instability, especially on lower-end GPUs. Still, we saw good improvements after tuning the models. The fine-tuned versions performed noticeably better at solving programming problems and showed stronger reasoning compared to their base versions. Our results suggest that even smaller models can be useful for code-related tasks if trained carefully. This makes it more practical to use such models in everyday scenarios where large-scale hardware isn’t available.