Traditional fire detection systems often suffer from high false detection rates, necessitating more accurate solutions, particularly in constrained environments. This study utilizes deep learning models, specifically the Google Vision Transformer, to develop an automatic fire hazard detection system. The model is trained on the RawFire dataset, combining TrainFire and TrainSmoke datasets from Kaggle, containing three classes: Fire, Smoke, and Normal. With 6060 training, 756 validation, and 759 testing images, the model is optimized over 10 epochs with specified parameters. Notably, it achieves remarkable performance metrics, including an accuracy, F1 score, precision, and recall of 0.99. Further experimentation focuses on determining the optimal number of attention heads, with the model demonstrating peak performance with 8 attention heads. This underscores the potential of fine-tuning transformer architectures for enhancing fire detection accuracy in real-world scenarios.