Small Language Models (SLMs) Can Still Pack a Punch: A survey

TL;DR

This survey analyzes Small Language Models (1–8B) as competitive alternatives to giant LLMs, arguing that data quality, training techniques, and architectural innovations can yield performance rivaling or surpassing larger models. It categorizes SLMs into task-agnostic, task-specific, and domain-specific and surveys advances in training methods (instruction tuning, distillation, CoT), data strategies, and efficient architectures like MoE and Hybrid State Space Models. A key concept is the effective size, showing several SLMs outperform larger counterparts on benchmarks despite far fewer parameters, driven by data quality and targeted training. The work also identifies practical implications for edge deployment, ongoing evaluation, and future directions, including revised scaling laws and broader benchmarking.

Abstract

As foundation AI models continue to increase in size, an important question arises - is massive scale the only path forward? This survey of about 160 papers presents a family of Small Language Models (SLMs) in the 1 to 8 billion parameter range that demonstrate smaller models can perform as well, or even outperform large models. We explore task agnostic, general purpose SLMs, task-specific SLMs and techniques to create SLMs that can guide the community to build models while balancing performance, efficiency, scalability and cost. Furthermore we define and characterize SLMs' effective sizes, representing increased capability with respect to LLMs.

Figures (2)

• Figure 1: Mind map of topics covered in the paper
• Figure 2: Equivalent sizes of SLMs based on performance benchmarks; more details in Table \ref{['tab:model-specs']}