<p>Wildfires have become a global threat in recent times. Not only do they cause the loss of forest resources and ecological imbalance, however, they also result in the loss of human lives. The economic impact is also substantial, with significant costs associated with property damage, firefighting efforts, healthcare, and loss of productivity. Furthermore, wildfires contribute to climate change by releasing large amounts of carbon dioxide and other greenhouse gases into the atmosphere. Early fire detection can save human lives and mitigate these extensive economic and environmental damages. This proposed work explores the possibility of detecting forest fires from images using a zero-shot approach. In recent times, several visual foundation models have been trained on massive datasets, enabling us to apply these models to a variety of tasks, including zero-shot classification, segmentation, and object detection. The zero-shot approach allows us to classify or detect objects in images without any prior training. Recently, Contrastive Language Image Pre-training (CLIP) models and Large Multimodal Models (LMM) have gained attention due to their excellent zero-shot classification and detection capabilities. With text prompts alone, images can be accurately classified. In this work, we utilize several CLIP models and LLaVA (Large Language Vision Assistant) to classify images as either containing fire or not. The effectiveness of these models is demonstrated by applying them to six wildfire datasets. Classification accuracy is computed, and a comparison of the results from both CLIP and LLaVA models is presented. Simulation results show that these models achieve accuracies beyond 95%, outperforming state-of-the-art methods in forest fire detection. To the best of the authors’ knowledge, this is the first implementation of CLIP and LLaVA models for forest fire detection.</p>