Recent developments in artificial intelligence have highlighted the capabilities of smaller language models (LLMs) in outperforming larger counterparts. A study indicates that a 1 billion parameter LLM can surpass a 405 billion parameter model on reasoning tasks, such as the MATH-500 benchmark, by utilizing compute-optimal Test-Time Scaling (TTS). The smaller models demonstrate a 14.1 times faster inference rate and require 256 times fewer floating point operations per second (FLOPS) compared to their larger counterparts. Furthermore, a 1.5 billion parameter model named DeepScaleR has reportedly outperformed OpenAI's O1-preview in complex mathematical reasoning, challenging the prevailing notion that larger models are inherently superior.