Baidu has unveiled an optical character recognition system it calls "Unlimited OCR," which can read dozens of document pages in a single pass. According to The Decoder, that is a notable jump over previous systems, which topped out at roughly ten pages before running into limits.
Optical character recognition, or OCR, is the technology that turns images of text — scanned contracts, forms, printed reports — into machine-readable words. The hard part is scale. Most models keep every page they have read in active memory, so processing more pages steadily consumes more resources until the system chokes.
Baidu's approach, as described by The Decoder, is to treat memory more like human forgetting. Rather than hoarding every detail of every page, the model uses a modified attention mechanism that keeps memory use flat no matter how many pages it works through. Attention is the part of a modern AI model that decides which pieces of information to focus on; by reshaping how it handles older pages, Baidu avoids the ballooning memory cost that caps rival systems.
The payoff is both practical and measurable. The Decoder reports that Unlimited OCR currently holds the top spot on the most important OCR benchmark — the standardized test researchers use to compare these systems head to head.
The source item does not detail how the model was trained, what languages it handles, or whether it is publicly available, so those questions remain open.
Why it matters: long documents — legal filings, medical records, financial statements — are exactly the material businesses most want to digitize, and a system that reads dozens of pages at once without runaway memory costs could make that far cheaper and more reliable.