Document extraction that runs where the documents already are
Vision-language models served locally, reading layout, tables and handwriting at 83 tokens per second. For documents that can't reach a cloud API, and volumes where per-page pricing breaks.
Some documents can't go to a cloud API. Contracts under privilege, records with identifiable data in them, anything a client has undertaken not to send to a third party. And separately, some volumes simply break per-page pricing.
Both problems have the same answer: run the model where the documents already are.
Qwen2-VL and Qwen2.5-VL, served locally on a single 24GB GPU through llama.cpp, measured at 83 tokens per second. They handle layout, tables and handwriting, which is the point — traditional OCR reads characters, and a vision-language model reads a page.
Nothing leaves the machine. No per-page fee, no upload, no third-party processor to add to a data map.
The measurement matters more than the number. Knowing the throughput a given model gets on a given card is the difference between quoting a document pipeline honestly and guessing at it.
Working on something in this territory? Tell us what you’re building