Abstract
The article explores the trade-off between efficiency and performance of small language models (SLMs) with fewer than 15 billion parameters, which represent a relevant alternative to resource-intensive large language models (LLMs). A comparison of modern SLMs, such as Llama 3.2 3B and Llama 3 8B, Gemma 2B and 9B, Qwen2.5 3B, 7B, and 14B, Phi-4 14B, and Mistral NeMo 12B, is conducted using standardized benchmarks (MMLU-PRO, GPQA, IFEval, MATH, BBH) to evaluate their capabilities in text generation, summarization, question answering, and logical reasoning. The results show that some SLMs demonstrate performance close to that of high-parameter models like GPT-4o, with significantly lower computational costs. The work highlights the potential of SLMs for creating more accessible and environmentally friendly solutions in the field of artificial intelligence, offering practical recommendations for researchers and developers.