Multimodal models
Models that read images, sound or video as well as text: GPT-4o, Gemini, LLaVA. 104 models on EveryLLM so far.
Multimodal models take in more than text. Most read images, such as photos, charts, handwriting and scanned documents; many also listen to speech, watch video or produce images and audio themselves. Inside, the other kinds of input are turned into tokens or vectors that the language model can attend to alongside words, so one network can answer a question about a picture as easily as one about a paragraph.
Joining vision and language
OpenAI's CLIP (January 2021) learned to match images with their captions, and DeepMind's Flamingo (April 2022) showed a language model answering questions about images after seeing a handful of examples. In 2023 the open LLaVA project linked a vision encoder to a language model with a small adapter, a cheap recipe that many open models still follow. GPT-4 accepted image inputs from its launch in March 2023.
Natively multimodal models
Google's Gemini, announced in December 2023, was trained from the start on text, images, audio and video together rather than having vision bolted on later. OpenAI's GPT-4o (May 2024) did the same for speech, answering spoken questions in about 320 milliseconds on average. Meta's Llama 4 (April 2025) was its first natively multimodal family. Some models, such as Google DeepMind's Gemini Robotics (2025), turn what they see into actions for robots.
Famous multimodal models
Still to be written: GPT-4o.