Method

AI OCR in your browser: no upload, no page limit

Every OCR tool on this site runs a neural network in your browser tab, on your graphics chip where the browser can use it. Nothing is uploaded, so there is nothing to meter: no daily allowance, no account, and no page cap on the fast reader; the Deep reader takes twenty pages a run, which is time on your machine, not a quota. This page says which model reads what, what each costs to load, what it was measured at, and where each stops.

Two readers

The fast reader is PaddleOCR's PP-OCRv6 — a text detector and a line recogniser, 6 MB of models on ONNX Runtime Web — behind Image to text, OCR a PDF and Make PDF searchable. It reads English and forty-six other Latin-script languages, Chinese and Japanese with one model, and Russian, Ukrainian and Belarusian with a second recogniser (PP-OCRv5, 8 MB) that loads when you choose it. About sixteen megabytes the first time — the runtime, the detector and the recogniser, compressed — then cached. Measured on an Apple M4 Mac at 200 dpi: 0.6 seconds an A4 page on WebGPU in Chrome and Safari; about two seconds on one processor thread in Chrome, six to eleven in Safari. Accuracy on a clean printed scan: 0.2 % of characters wrong in English, 0.7 % in Portuguese, 0.6 % in Russian.

The Deep reader is GLM-OCR — a 0.9-billion-parameter vision-language model from Zhipu AI, MIT-licensed, run with Hugging Face's browser library, transformers.js — behind the Pro door on OCR a PDF and Image to text, and as the whole of Image to Excel. It reads a page as a page: paragraphs, headings, a formula as LaTeX, a table as a table. It costs about 660 MB, downloaded once from Hugging Face's servers and kept in this browser, and about fifteen to twenty seconds a page on a recent laptop's graphics card — measured 18 seconds in Chrome and 15 in Safari on the same Mac. It needs WebGPU — the browser's way of using the graphics chip; Chrome, Edge and Safari 26 have it — on a laptop or desktop with at least 8 GB of memory, and refuses a phone or tablet before any download.

What "unlimited" means here, and what it does not

Online OCR services upload your file and cap what they will read: a few pages an hour, one file at a time, a monthly allowance behind a login. Here there is no server, so there is nothing to cap — a forty-page scan costs you forty times a page's reading time on your own machine and nothing else. That is the whole of the promise. It does not mean fast: the Deep reader takes a quarter of a minute a page, and the page says so before you start.

What each refuses

The fast reader refuses a PDF that already carries text, because reading what is there beats guessing at it, and it does not promise handwriting, which it was never measured on. The Deep reader refuses a phone or tablet, a device with under 8 GB of memory, a browser without WebGPU, a graphics chip that cannot hold a one-gigabyte buffer or compute in 16-bit floats (the weights are stored that way), a document over twenty pages, and a picture whose "table" answer has no rows and columns in it. Every refusal names itself and the tool that can help instead — for the Deep reader's refusals, the fast reader on the same page.

Where the models come from

The fast reader's models are PaddlePaddle's own release archives, served from this site. The Deep reader's weights are the onnx-community conversion of GLM-OCR, fetched from Hugging Face at one pinned version — aea46198f09e — and checked against the digest of every 64 MB piece before they are used; a piece that does not match is refused and nothing of it is kept. Every library is on the licenses page with its licence. What leaves this device, and to whom, is on the method page and the privacy page.

About "Pro"

The Pro door shows a price and then says the truth: Deep OCR is free during the preview, with no card and no account. The price is there to learn whether anyone would pay for a reader of this weight before one is charged for. What "Pro" records is which of two prices you were shown and whether the preview was started, loaded, finished or refused — six event names, nothing attached; the privacy page lists them.