Turn a page into markdown, on your own GPU
Drop a scan, a photo of a page or a PDF. A small OCR model finds the text lines, reads them and puts them in
reading order, with headings, lists and tables, all inside this tab: the page is not uploaded anywhere. It is
Indra OCR by Datalab, its weights stored as 16-bit floats, which score the
same as the originals: the first visit downloads about 4 MB. It needs a browser with WebGPU; it was tested in Chrome on a Mac.
How it works: a detector finds the text lines. A recognizer reads each line; the first one reads Latin, Greek, Cyrillic,
Devanagari and Bengali, and three more are downloaded (2 to 9 MB each, with their word models) the first time a line needs them: Chinese,
Japanese and Korean; Arabic and Hebrew; and other scripts such as Thai, Ethiopic and the South Indian scripts. A line the
recognizer is unsure of is read again with a word model (an n-gram table of which characters usually follow which) that
breaks close calls; the Latin one (2.5 MB) is downloaded the first time a line needs it. A small transformer then puts the lines in reading order and labels headings, lists and tables.
Everything runs on WebGPU, in code ported from Datalab's indra-ocr
Python package: given the same weights, the two produce the same text and markdown on the pages compared.
Write-up: indra-browser .
The model weights are Datalab's, under a modified OpenRAIL-M license (full license text ):
free for research and personal use; companies with more than $20M in revenue or funding, or that compete with Datalab,
need a commercial license from Datalab. The weights and word models served here are Datalab's, rounded to fewer bits
(weights to 16-bit floats, word-model values to 2 bits); nothing else in them is changed. Build .