Recent developments in the field of large language models (LLMs) highlight a trend towards optimizing smaller models for enhanced performance. Researchers from Hugging Face have demonstrated that their 3 billion parameter LLaMA model can outperform larger models, such as the 70 billion parameter variant, on mathematical tasks by utilizing test-time compute scaling techniques. This approach allows models to process information more effectively during problem-solving. Additionally, Meta has introduced a new technique that enables its 1 billion parameter LLaMA to surpass the performance of its 8 billion counterpart in similar tasks. The advancements in test-time compute scaling have been recognized by various organizations, including Google DeepMind, which has explored methods to optimize performance on challenging tasks. Furthermore, AI2's OLMo 2 has set a new benchmark by outperforming Meta's LLaMA, demonstrating the potential of smaller models trained on extensive datasets. These innovations reflect a broader shift in AI development towards more efficient and accessible LLMs, with Hugging Face's open-source approach leading the way in this evolving landscape.