Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Baidu Releases Unlimited-OCR 2: Model Reads 5026 Pages in a Single Pass

OCR, or optical character recognition, extracts text from scans and photos. Long documents are usually split page by page, so a table that continues onto the next page falls apart. Unlimited-OCR 2 from Baidu reads PDFs, scans and photos as a whole and remembers previous pages. In one pass, it processed all seven Harry Potter books, more than 3600 pages, and in another test a 5026-page document. For digitizing archives, that means less manual stitching.

AI-processed from TG: @notboring_tech (Not Boring Tech); edited by Hamidun News
Baidu Releases Unlimited-OCR 2: Model Reads 5026 Pages in a Single Pass
Source: TG: @notboring_tech (Not Boring Tech). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @notboring_tech (Not Boring Tech) ↗
Editor-in-Chief
Key metric #1Extracts text, tables and formulas from PDFs, scans and photos
Key metric #2Does not reread the entire document from scratch, but pulls the relevant fragments
The code is open source and freethe baidu/Unlimited-OCR repository on GitHub

Imagine you need to turn an archive of several thousand pages into text. Optical character recognition (OCR, the thing that "photographs a page and gives you letters") usually works on each page separately. As a result, a table that starts on one page and continues on the next becomes two unrelated chunks, and putting the document back together falls to people or scripts.

Unlimited-OCR 2, released as open source in Baidu's GitHub repository, tries to solve this problem: according to the description, it reads the document as a whole and keeps track of what appeared on previous pages. In a demo, the model processed all seven Harry Potter books, more than 3600 pages, in a single pass, and in another test it handled a 5026-page document.

What Unlimited-OCR 2 can do

The model is distributed free of charge with open-source code. According to the description, it extracts not only plain text but also tables and formulas from PDFs, scans and ordinary photographs. This matters: tables and math are the elements that most often break during recognition, turning into a mess of symbols.

The main claimed feature is the preservation of context across pages. If a table continues onto the next spread, the model understands that it is one and the same table, not two different ones. Take a 40-page accounting report: ordinary OCR would output forty separate fragments, while here the result should be a coherent structure.

How the model handles thousands of pages

The longer the document, the more memory is needed to hold it "in mind." The authors claim that Unlimited-OCR 2 saves memory: it does not reread the entire document from scratch, but accesses only the fragments it needs. A rough analogy is a reader who does not flip through a book from the first page each time they need a quote, but opens a bookmark.

The figures from the demonstrations look impressive. All seven Harry Potter books amount to more than 3600 pages, and the model processed them in one pass. In another test, it handled a document of 5026 pages. Volumes like these usually have to be split into many separate runs and the results stitched together afterward; here, according to the authors' design, that step is not needed.

Who will benefit

The first obvious candidates are:

  • libraries and archives digitizing books and old periodical bound volumes;
  • research teams with thousands of papers containing formulas and tables;
  • legal and finance departments with thick contracts and reports;
  • developers of RAG systems (search for answers over your own documents), who need clean text with preserved structure.

For the last group the benefit is especially tangible: the quality of a RAG answer depends directly on how accurately the document was converted to text. If a table at the input falls into two unrelated fragments, no language model will later reconstruct the correct figures.

What to check before deploying

So far this is an announcement, and the numbers come from the project description. The reports cite page counts, but not recognition accuracy in percentages, speed on specific hardware, or VRAM requirements. The Harry Potter test consists mainly of clean printed text; how the model performs on skewed scans, handwritten annotations, old newspapers or tables with complex layouts will only become clear in practice.

A reasonable plan is as follows. Pick a hundred of the trickiest pages from your archive: tables that continue across pages, formulas, photos taken in poor lighting. Run them through Unlimited-OCR 2 and through your current solution. Compare the results by hand: how many tables were assembled correctly, how many formulas were lost, and how much memory and time were used. If the gain is noticeable, it makes sense to build a digitization pipeline around the model. Open code lets you run it on your own infrastructure without sending confidential documents to someone else's cloud.

Overall, the release reflects a broader trend: document recognition is ceasing to be a "page-by-page" task and becoming a "whole-document" task. For practitioners this is good news, and the claims can be verified independently, since the code is in Baidu's public repository.

Story

Key Questions Answered

What Unlimited-OCR 2 can do?

The model is distributed free of charge with open-source code. According to the description, it extracts not only plain text but also tables and formulas from PDFs, scans and ordinary photographs. This matters: tables and math are the elements that most often break during recognition, turning into a mess of symbols. The main claimed feature is the preservation of context across pages. If a table continues onto the next spread, the model understands that it is one and the same table, not two different ones.

How the model handles thousands of pages?

The longer the document, the more memory is needed to hold it "in mind." The authors claim that Unlimited-OCR 2 saves memory: it does not reread the entire document from scratch, but accesses only the fragments it needs. A rough analogy is a reader who does not flip through a book from the first page each time they need a quote, but opens a bookmark. The figures from the demonstrations look impressive. All seven Harry Potter books amount to more than 3600 pages, and the model processed them in one pass.

Who will benefit?

The first obvious candidates are: libraries and archives digitizing books and old periodical bound volumes; research teams with thousands of papers containing formulas and tables; legal and finance departments with thick contracts and reports; developers of RAG systems (search for answers over your own documents), who need clean text with preserved structure. For the last group the benefit is especially tangible: the quality of a RAG answer depends directly on how accurately the document was converted to text.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…