Abstract: Karst carbon sinks play a key role in narrowing the global carbon-balance gap. Reconstructing DIC (dissolved inorganic carbon)/DOC (dissolved organic carbon) concentrations with machine learning and couple them with the hydrochemistry-runoff method to more precisely quantify basin-scale karst carbon-sink flux. This study focuses on the Lijiang River Basin in Guangxi Zhuang Autonomous Region, China. Based on two years of field measurements (September 2015 to September 2017), we compared six machine-learning algorithms and selected random forest to build DIC and DOC concentration models (R² reached 0.93 and 0.68, respectively). We then predicted monthly DIC/DOC in 2017 and, with the hydrochemistry-runoff method, estimated a carbonate carbon sink flux of 2.42 × 10⁵ tCO₂/(km²·yr) and an intensity of 43.41 tCO₂/(km²·yr). The concentration prediction results are consistent with previous studies, supporting the reliability of both the model and the accounting. This study provides methodological innovation and a scientific basis for karst carbon-sink estimation, and offers a reference for application in sparsely monitored karst basins.