Abstract: The energy efficiency and carbon emission intensity of gas-fired boilers are critical metrics for achieving car-bon peaking and carbon neutrality goals. Traditional testing methods are time-consuming and hinder rapid as-sessment. Based on field test data from more than 200 non-condensing gas-fired boilers, six key characteristic parameters are selected using Pearson correlation analysis, including methane content, carbon dioxide content, feedwater flow rate, inlet cold air temperature, exhaust gas temperature, and heat loss ratio. Taking thermal efficiency and direct carbon emission intensity as the prediction targets, the predictive performances of five models, namely Neural Network, Random Forest, XGBoost, LightGBM, and Support Vector Regression, are compared. Five-fold cross-validation results show that the XGBoost model achieves the best performance, with R2 values of 0.958 and 0.892 for thermal efficiency and carbon emission intensity, respectively. The SHAP method is employed to reveal the direction and magnitude of the influence of each feature on the two targets. The results indicate that exhaust gas temperature is the most critical affecting factor. Furthermore, among the fuel characteristics of natural gas, methane content exhibits a negative correlation with direct carbon emission intensity, while carbon dioxide content shows a positive correlation. Based on the optimal model, a prediction platform is constructed, enabling real-time input of operating parameters and visual output of both predicted targets. This study provides an efficient and interpretable predictive model that achieves acceptable accuracy with a limited set of measured parameters, which offers the potential for developing low-cost, robust online monitoring tools for low‑carbon operation of gas‑fired boilers.