Vision-Language Model (VLM)

What is a vision-language model?

Old-school AI handled text and images separately — one model read words, another recognized pictures. A vision-language model (VLM) joins the two together: it can both "see" images and "speak" words, and it connects the two streams of information. Show it a photo and it tells you what's happening; give it a word problem with a diagram, and it reads the picture before answering.

How is it different from plain image recognition?

Image classification just gives a label
A classic vision model says "this is a cat" and stops there.
A VLM can talk and reason
A VLM answers "what is the cat doing", "why is there a shadow here", and can even read the text, charts and diagrams inside an image.

How does it work, roughly?

Turn the image into "image language"
A vision encoder converts the picture into a sequence of numeric features — a kind of description code for the image.
Plug the image into a language model
That code is fed to a large model together with the text, aligning words and images in one space so the model can talk and reason about what it sees.

What's it good for?

From "snap a photo and it tells you what the dish is" to "upload a screenshot and it writes your copy", and in accessibility, describing pictures for people who can't see them — VLMs stretch AI's perception from pure text out into the real world.

Bottom line: a vision-language model gives a language model eyes, so it can see images and tell you what's in them.