A modality is a kind of input or output.
Text, still images, audio, and video carry information in different forms. A multimodal system can use more than one of them in a task—for example, answer a question about a photograph, turn a voice recording into text, or produce spoken audio from a written response.
The word describes the range of information a system handles. It does not say how the system is built, how well it performs, or whether it can create every format it can accept. Image understanding, image generation, speech recognition, and speech synthesis are distinct capabilities that need their own evidence.
A product can offer a unified interface while routing work through several components: an audio recognizer, a language model, retrieval, and a speech generator. Other architectures train a model to work directly across multiple modalities. The interface alone does not reveal which design is underneath.
These are design choices, not a ranking. A pipeline can make individual transformations easier to inspect; a joint model may handle relationships that a transcript or caption would discard. Choose by the task and check the evidence each route exposes.