A base model is a large language model that has been trained on vast text data to predict the next token but has not yet been fine-tuned for specific tasks or aligned with human preferences. It is raw talent: brilliant at completing text but not yet trained to follow instructions politely or refuse harmful requests. Base models serve as the starting point for creating chatbots and assistants through further training stages like instruction tuning and RLHF.