Soft prompts have three fundamental limitations: they are deeply bound to a specific model architecture and cannot be transferred across models; the continuous vectors cannot be directly understood by humans; and they require access to the model's internal embedding layer during inference, making them incompatible with commercial models that provide only API access.