Skip to content
agentscamp
Term · Term

VLM (Vision-Language Model)

A VLM jointly understands images and text — reading documents, screenshots, charts, and photos and reasoning about them in language.

Updated Jun 11, 2026
vlmvisionmultimodalocr

A vision-language model (VLM) is a model that takes images alongside text and reasons over both — describing a photo, extracting a table from a scanned invoice, reading a dashboard screenshot, or explaining a chart.

Architecturally, a vision encoder turns images into tokens the language model attends to natively, so the LLM's reasoning applies directly to visual content. That collapsed what used to be pipelines: OCR + layout analysis + parsing became one call to a model that reads the page — the shift covered in Using VLMs for OCR, Documents, and Video, and packaged for extraction work in the multimodal-document-extractor skill.

Frontier APIs are all VLMs now, and open-weight families like Qwen3-VL brought the capability to self-hosting. Beyond documents, VLMs are the perception layer of computer-use agents (reading UIs to act on them) and of coding agents verifying their own frontend work from screenshots. The practical craft: manage image resolution deliberately — it's both the accuracy ceiling for small text and the token bill.

Frequently asked questions

Do VLMs replace OCR?
For most document understanding, effectively yes. Classic OCR outputs characters and leaves structure to you; a VLM reads the page like a person — tables, layout, handwriting, checkboxes, the figure the text refers to — and can return structured data directly. Dedicated OCR still wins on raw character accuracy for clean, high-volume scanning at minimal cost.
What are VLMs' weak spots?
Precise counting and measurement, dense small text at low resolution, exact spatial coordinates, and hallucinated detail when an image is ambiguous — the model fills gaps plausibly, like any LLM. Resolution settings matter more than people expect: token cost scales with image size, and downscaling silently destroys small text.

Related