Software Generation for Embedded Systems with Low-Cost LLMs: An Energy Evaluation Discussion

Abstract Energy efficiency has become a central concern in the deployment of AI systems. While most research targets hardware-level acceleration to reduce energy consumption, software-based strategies remain underexplored or often sacrifice performance. In this work, we present a software-oriented approach for embedded code generation using a multi-agent pipeline composed of small, locally executed LLMs. Our architecture divides the generation process into reasoning stages, aiming to reduce resource demands without relying on cloud services. We evaluated our system using real-world measurements of energy and memory consumption, comparing multiple model configurations and execution scenarios. The results suggest that running small LLMs locally is a feasible path for software generation under constrained resources, contributing to the broader discussion on low-energy AI without specialized hardware.