Sweetness is one of the critical quality indicators in cherry tomatoes. Traditionally, measuring sweetness requires a destructive method: the fruit must be cut open, the juice extracted, and then analyzed using a testing device. However, destructive analysis and high-resolution spectroscopy have notable drawbacks, such as high investment costs and non-portable equipment, making them unsuitable for on-site measurements or industrial applications. This study employed a combination of spectroscopy with wavelengths in the visible and near-infrared (VIS-NIR) range and an AI-driven method to predict the sweetness of whole cherry tomatoes. Before modeling, the spectral data were preprocessed using the SNV method. Multiple Linear Regression, Support Vector Regression, and Gaussian Process Regression were applied to model the complex relationship between spectral features and the sweetness values obtained through destructive testing. The GPR model achieved the highest performance for the raw spectral dataset, with a 0.86 coefficient of determination (R2) and a 0.33 brix Root Mean Square Error (RMSE). In contrast, SVR combined with SNV preprocessing yielded the highest accuracy, with an R2 of 0.88 and an RMSE of 0.30 brix. The proposed system demonstrates strong potential for non-destructive sweetness assessment in cherry tomatoes, particularly in on-site settings and industrial environments where rapid, cost-effective, and portable analysis is essential.