What is a Multimodal AI?

Early AI systems were largely unimodal: a language model read and wrote text, a vision model classified images, a speech model transcribed audio, each trained and deployed separately. Multimodal models instead learn a shared representation space where, for example, an image and its text description map to related internal representations, letting a single model answer a question about a photo, describe a chart, or generate an image from a text prompt. This matters practically because most real-world tasks aren't confined to one modality, reading a scanned invoice, watching a video for a specific event, or reviewing a screenshot alongside a bug report all require combining visual and textual understanding. Multimodal models can also chain modalities: transcribing audio to text, reasoning over that text, and producing a spoken response, unifying what used to require several separate specialized systems glued together.