Energy-Efficient Large Language Models

Abstract

The ecological and economic burden of large pre-trained language models (LLMs) is now on par with mid-sized manufacturing facilities. Recent work has shown careful combinations of sparse architectures, parameter-efficient fine-tuning, and aggressive low-precision inference techniques can reduce total energy use by 35 – 73% maintaining task accuracy. In this evidence synthesis, we review ten significant studies published from 2024 to 2025, which prioritise energy efficiency across the entire LLM lifecycle. We categorise the studies within three layers: architecture/pre-training, fine-tuning, and inference - synthesising benefits; practical trade-offs; benchmarking gaps; and open research questions. Finally, we provide a roadmap to achieve sub-50 Wh chatbot sessions through hardware-software co-design and standardised reporting of the entire lifecycle.