Understanding and predicting the temporal popularity of social media content is critical for digital marketing and recommendation systems. Existing methods often rely on static feature concatenation, struggling to capture complex multimodal interactions and strict sequential dependencies in long-term popularity evolution, frequently underestimating viral content. To address these limitations, we propose a MultiModal Autoregressive Transformer Framework (MMATF). Our framework extracts and refines key representations from video covers, texts, category tags, and numerical statistics. A dynamic modality interaction module adaptively fuses these inputs to capture evolving content dynamics, establishing a solid foundation for temporal relevance. Furthermore, an autoregressive temporal encoding module strictly prevents future information leakage. Extensive experiments on SMTPD and ICIP datasets demonstrate MMATF significantly outperforms advanced baselines across 7-, 14-, and 30-day horizons. Specifically, compared to the strongest baseline, our model reduces Mean Absolute Error by 3.8% and 4.3%, and improves Spearman’s Rank Correlation by 0.3% and 0.4% for 30-day predictions on SMTPD and ICIP respectively, proving consistent long-term forecasting accuracy.