Inner speech (IS), or imagined speech without overt articulation, is a promising target for brain-computer interfaces (BCIs) aimed at restoring communication in individuals with severe speech impairments, such as locked-in syndrome. Foundation models (FMs), typically trained using self-supervised learning (SSL) on large-scale datasets, offer new opportunities for learning transferable and robust representations from neural signals. This mini review provides an overview of FM-based approaches for IS decoding using non-invasive neuroimaging modalities, including functional magnetic resonance imaging, electroencephalography, magnetoencephalography, and functional near-infrared spectroscopy, highlighting architectural trends, pretraining strategies, and model adaptation techniques. We discuss how recent models move beyond task-specific classification toward scalable representation learning and semantic-level decoding. Despite these advances, several challenges remain, including the weak, noisy, and non-stationary nature of neural signals, variability in data acquisition, and limitations in dataset scale, standardization, computational resources, interpretability, and evaluation metrics. Ethical and privacy considerations are also critical. Overall, FMs provide a promising paradigm for non-invasive IS decoding, addressing neurophysiological, methodological, and ethical challenges is essential for developing scalable and reliable BCI systems.