Accurate short-term electricity demand forecasting is essential for effective energy planning, particularly in small communities with limited infrastructure and scarce historical data. In such settings, conventional forecasting techniques that require extensive training are often impractical due to high computational demands and insufficient data availability. This study addresses this limitation by evaluating the zero-shot performance of two pre-trained time series forecasting models: Lag-Llama and TimesFM 2.0. Real-world electricity consumption data were used, collected from three households in a Mexican community at 15-minute, hourly, and daily intervals over a one-year period. A standardized evaluation framework based on a sliding window approach and Root Mean Squared Error was employed. Results indicate that TimesFM 2.0 outperforms Lag-Llama in high-frequency forecasting, while Lag-Llama yields competitive or superior performance in certain hourly-resolution scenarios. These findings demonstrate the practical viability of zero-shot forecasting with foundation models in data-constrained environments, offering a scalable alternative to traditional model training.