The field of time series forecasting is moving towards the integration of multimodal data, including numerical and textual information. Recent developments have focused on incorporating large language models (LLMs) and state space models to improve forecasting accuracy and provide human-readable summaries of forecasts. The use of hierarchical attention mechanisms and mixture of experts frameworks has also shown promising results in capturing complex temporal dependencies and improving performance on benchmark datasets. Noteworthy papers include LLM-Integrated Bayesian State Space Models, which introduces a novel probabilistic framework for multimodal temporal forecasting, and Xihe, which proposes a scalable zero-shot time series learner via hierarchical interleaved block attention. Additionally, MAP4TS and EMTSF have demonstrated the effectiveness of multi-aspect prompting frameworks and mixture of experts models in time series forecasting. TempoPFN has also shown competitive performance in zero-shot time series forecasting using synthetic pre-training of linear RNNs. Overall, these developments highlight the potential of multimodal approaches and advanced architectures in improving the accuracy and interpretability of time series forecasts.