What is a vision language model?
A language model that can also look at images. The reason a scanned page, a table or a handwritten note can get an answer instead of a conversion error.
A vision language model is a language model that can also look at images. Instead of converting a page to text first, it reads the photograph of the page directly, so layouts, tables and handwriting that defeat plain OCR stay reachable. It runs locally on modest hardware in quantized form. Its reading is still a model's reading: grounded tools attach what it sees to citations and check the claims.
The model that reads the page itself.
Plain OCR is a specialist: a narrow engine that turns marks into characters, and Windows ships one, which is how scanned pages become searchable. A vision language model takes the wider route. It is a language model with a second front end, a vision encoder that turns an image into the same kind of token stream the text path uses, so reading a page and answering about it happen in one model instead of two. That is what people mean by a model with eyes: not a camera bolted on, but a second way into the same reading machinery.
What it sees that OCR misses.
Because the model reads the whole page image, it keeps what a character-by-character conversion throws away: table structure, column order, figures, marginalia, and the visual context that separates a heading from a footnote. Handwriting is the sharpest test, and what local vision models manage there is real but bounded, good on neat script and strained by hurried cursive. The trade is cost: an image becomes a block of extra tokens, so vision runs warmer and slower than text alone, which is why quantized vision models are the ones that fit on ordinary hardware.
Seeing is not checking.
A model that can look at a page can still misread it, confidently. It can transcribe a line wrong, describe a figure that is not there, or answer from what pages like it usually contain rather than what this page says. The remedy is the same one text answers need: citations that point at the actual page and a verification pass against what was read. That is why grounded tools treat the vision model as a reader, not an oracle, and why the structural fixes for fabrication apply with eyes exactly as they do without them.