Strategies for computational efficiency in small language models

Abstract The proliferation of Large Language Models (LLMs) is often constrained by their significant computational and memory requirements, limiting their deployment to large data centers. Small Language Models (SLMs) offer a feasible solution for on-device applications; yet their efficiency requires optimization to operate well on resource-constrained hardware. This study looks at ways to make SLMs more efficient at using computers. The effects of two primary methods were compared: post-training optimization and architectural innovation through quantitative and observational study. Using a standardized suite of benchmarks measuring accuracy, reasoning, and inference performance, a baseline is established with state-of-the-art SLMs like Phi-3 and Llama 3. Post-training techniques were evaluated, including 4-bit quantization (GPTQ) and knowledge distillation from a superior teacher model. Finally, these optimized Transformers were compared against a custom-trained, non-Transformer model based on the Mamba (State-Space Model) architecture. Results show that 4-bit quantization is the most effective compression strategy among those evaluated. It reduces peak inference memory footprint by 71%, increases throughput by 83%, and does so with minimal accuracy degradation. Within the controlled experimental space evaluated in this study, the 4-bit quantized Phi-3-mini model occupies a Pareto-optimal position in memory-normalized accuracy. Focused skill growth works best with knowledge distillation. However, new designs like Mamba offer a different trade-off by being the best at streaming jobs’ raw output. It was found that improving current Transformer-based SLMs through quantization is the best way to use them for general purposes. However, customized designs and distillation work better for specific, high-performance uses. This research offers a definitive framework and pragmatic recommendations for advancing the next generation of robust, efficient, and accessible language models.