WiFi Channel State Information (CSI) has emerged as a promising modality for device-free human sensing tasks, such as human activity recognition (HAR) and identity recognition (Human ID), due to its ubiquity, low cost, and non-intrusive nature. Traditional model-driven approaches, based on physical propagation theories (e.g., Fresnel zone models), offer interpretable features but lack the capacity to capture complex semantic patterns. Learning-based methods improve performance by integrating domain knowledge with deep neural architectures, yet they remain limited under task shifts, scarce data, and unseen classes. In this work, we explore the use of large pre-trained time-series foundation models (TSFMs)—originally developed for domains such as weather forecasting and electric signal analysis—for CSI-based human sensing. These models exhibit strong cross-task generalization, jointly model frequency–temporal dependencies, and are inherently suited for few-shot learning and multimodal fusion via a modular pretraining–finetuning paradigm. We further investigate parameter-efficient fine-tuning (PEFT) strategies, including LoRA and prompting, to adapt state-of-the-art TSFMs to CSI data. Experimental results on HAR and Human ID tasks demonstrate that our approach consistently outperforms existing baselines across various settings, providing a scalable, data-efficient, and generalizable framework for wireless sensing in AIoT applications.