Carbon monoxide (CO) is a hazardous airborne pollutant and a very toxic gas that poses serious health risks and contributes to greenhouse effects in the atmosphere. Accurate prediction its concentration is crucial to minimize human exposure, reduce global warming, support informed public health initiatives, and ensure compliance to strict environmental standards. Due to its significant swings in concentration and frequent presence within urban and factory zones, CO levels in the atmosphere are known to fluctuate rapidly under the influence of environmental variables like traffic emissions, climatic conditions, and industrial activities. Predictive modeling, particularly when utilizing deep learning models paired with quality real-time data, provides means for a precise estimation of CO concentrations in the atmosphere with high accuracy. This study applies recurrent neural networks (RNNs) to address this problem, recognizing that optimal forecasting performance drastically depends on the careful optimization of model hyperparameters. For mitigating this challenge, this paper suggests a tailored variant of the powerful and widely known firefly algorithm, designed specifically to improve the tuning procedure of RNNs. The capabilities of the suggested system were evaluated utilizing actual air quality monitoring data, yielding encouraging outcomes in predicting atmospheric CO levels.