Zero-shot inference in the CLIP paper refers to the model's ability to directly perform image classification using only natural language prompts, without any task-specific fine-tuning. CLIP pretrains an image encoder and a text encoder through contrastive learning, mapping images and text into a shared multimodal embedding space. At inference time, for a given image, the model generates embeddings for a set of text descriptions, computes the cosine similarity between the image embedding and each text embedding, and selects the text with the highest similarity as the predicted class. This approach allows CLIP to flexibly generalize to unseen categories, achieving cross-task transfer using only pretrained knowledge while avoiding the reliance on large amounts of labeled data that traditional models require.