The field of time series forecasting is witnessing a significant shift towards leveraging multimodal interactions, particularly with the integration of Large Language Models (LLMs) and semantic-spectral knowledge distillation. This direction is yielding innovative results, where frameworks are being designed to facilitate progressive cross-modality interaction, enabling textual information to effectively support temporal prediction. Furthermore, the application of foundation models, especially time-series foundation models (TSFMs), is showing strong performance on various tasks, including classification, regression, and imputation, with toolkits emerging to standardize and modularize their construction and fine-tuning. The emphasis on semantic abstractions and the utilization of dynamic semantic mechanisms to guide LLMs for time series forecasting is also a notable trend, leading to superior generalization capabilities in both zero-shot and few-shot settings. Noteworthy papers in this regard include FiCoTS, which introduces a fine-to-coarse framework for multimodal time series forecasting, and S^2-KD, a novel framework unifying semantic priors with spectral representations for distillation, enabling lightweight student models to make semantically coherent predictions. Additionally, STELLA guides LLMs with semantic abstractions for enhanced forecasting performance, and FMTK provides a modular toolkit for composing time series foundation model pipelines. These developments highlight the field's movement towards more integrated, efficient, and interpretable forecasting solutions.