The present disclosure provides a method, an Internet of Things system and medium for early warning smart gas harmful components. The method comprises: obtaining composition information of a gas and use information of a user; determining a first generation rate of the harmful components based on the composition information; determining a second generation rate of the harmful components through a generation rate prediction model based on the composition information and the use information, wherein the generation rate prediction model is a machine learning model and obtained by training, wherein a training sample includes historical use information of the user and historical composition information, and a label includes a second generation rate corresponding to the historical use information of the user and the historical composition information; determining a generation rate of the harmful components based on the first generation rate and the second generation rate; and generating warning information in response to the generation rate of the harmful components being greater than a generation rate threshold.