Multimodal AI refers to models that can process and generate multiple data types (text, images, audio, video). Two main architectural approaches exist: modular feature-level fusion, where a separate vision encoder extracts features and passes them to an LLM (cheaper but lossy), and native multimodality, where all data types are tokenized into a shared vector space enabling simultaneous reasoning across modalities. Native multimodality allows the model to attend to text and images together rather than pre-processing them separately. For video, newer models use spatiotemporal patches (3D cubes) that encode motion directly into tokens rather than sampling discrete frames. With a shared vector space, models can also perform any-to-any generation, accepting and producing any combination of modalities coherently.