This study developed and validated a deep-learning model for predicting Ki-67 expression in meningiomas using baseline MRI scans. A total of 1,239 patients from three medical centers were retrospectively analyzed, forming training, internal validation, and two external validation cohorts. The model employed a multi-modal representation learning framework incorporating MRI features, radiological characteristics, and radiomics data. Model performance was assessed through area under the curve (AUC) metrics and compared against traditional logistic regression and existing machine learning models. The deep learning model demonstrated superior predictive accuracy, achieving an AUC of 0.797 in internal testing and 0.808 in external validation. Additionally, the model was tested for its ability to predict tumor growth over three and five years, yielding AUCs of 0.756 and 0.727, respectively. Kaplan-Meier survival analysis confirmed a statistically significant association between model predictions and tumor progression. Importantly, intra-tumoral necrosis was identified as the strongest independent predictor of high Ki-67 expression (OR = 4.048, 95% CI: 2.878–5.694, p < 0.001). The model outperformed conventional clinical assessment methods and previous machine learning approaches, demonstrating improved generalization across datasets.