Home
>
PDF Tips
>
Baidu OCR in 2026: What's New in Unlimited OCR 2, How to Use It?
Baidu OCR in 2026: What's New in Unlimited OCR 2, How to Use It?
- 1 Baidu OCR contains three different products. A paid cloud API, PaddleOCR for open-source and offline work, and Unlimited OCR for whole documents in one pass. Picking the wrong one is the main reason OCR feels unreliable.
- 2 With Hugging Face space you can try Unlimited OCR 2 online free without installing model inference and deployment frameworks. You can also jump to read detailed steps guide.
- 3 The October 27 update changed the engine, not the size. Unlimited OCR 2 keeps the same 3B parameters and MIT licence as version 1, but swaps the attention mechanism.
- 4 Free is worth planning around. The cloud API gives 1,000 calls a month, PaddleOCR is free under Apache 2.0, and the Unlimited OCR weights cost nothing at any volume.
- 5 Privacy depends on the door you come in through. The open source model keeps your pages on your own machine, while the cloud API stores data in China and carries no contractual promise that your images are never used for training.
The updated code for Baidu Unlimited OCR 2 is now available on GitHub, but for most non-technical users, getting started remains the hardest part. Compared with earlier OCR releases, the new version delivers clear improvements in recognition accuracy, multilingual support, and layout analysis. Yet "open source" often means a barrier rather than freedom for everyday users.
This article did two things: it breaks down what actually changed versus historical versions, and provides a step-by-step guide to trying it online for free — no command line, no setup, just your browser. You don't need to configure anything arduously to test the new model. Just keep reading.
Part 1. What Is Baidu OCR? Three Product Lines, One Brand
Baidu ships three OCR model lines under one brand, and they do not compete. Most people who pick the wrong one end up paying for a problem they could have solved for free.
The Three Product Lines Compared
Please swipe to view
- Want the easiest, just-use-it-online option? Go with Baidu Cloud. Sign up, upload an image, and it reads it. Good for occasional use or building into a product.
- Want free, self-controlled, no-internet use? To run recognition on your own machine, Baidu PaddleOCR is the way to go. You'll need to install it yourself (a bit of setup), but it costs nothing.
- Need to read a whole book or a very long document? Go with Unlimited OCR. It's built to "read a whole file in one pass." The model is free, but the online service isn't fully open yet.
Part 2.What Changed in Unlimited OCR 2 on October 27, 2026
Baidu's own demo is hard to forget: Unlimited OCR 2 parsed all seven Harry Potter books, 3,623 pages, in one request, finishing in 90 minutes at about 1.5 seconds a page. The first generation, released on June 22, 2026, topped out near 40 pages. Version 2 published as what Baidu said moves the ceiling to effectively unlimited by narrowing attention to the pages that matter for the line it is writing.
Unlimited OCR 1 vs Unlimited OCR 2 Capability Comparison
Please swipe to view
Part 3. What Does Baidu Unlimited OCR Do?
In plain terms, Baidu Unlimited OCR reads documents. The output matters more: not a wall of text, but a structured document with headings, tables, formulas and reading order intact.
- Reads images, scans, screenshots and full PDFs, then returns Markdown or JSON
- Keeps structure, so headings stay headings, tables stay tables, formulas become LaTeX
- Merges tables split by a page break and rejoins cut paragraphs
- Reports each block's type and bounding box, so text maps back to the source
In one sentence: the model attends to the current page and holds only a small window of what it just wrote, so a 500-page document does not cost 500 pages of memory.
Supported Platforms and Product Formats
- Online demo, a Hugging Face Space where you drop in an image and watch Markdown appear
- Local open source on Windows, macOS or Linux, via Transformers or an SGLang server, MIT licensed, 6.7 GB of weights
- Baidu Cloud API, the Document Parsing (Unlimited-OCR) endpoint, returning Markdown plus a JSON
- Private deployment on your own servers, for organisations barred from a public cloud
- Mobile, through the Baidu app's camera recognition
Free Baidu OCR Options Compared
Please swipe to view
Part 4. How to Use Baidu Unlimited OCR
There are four sensible routes. Baidu Unlimited OCR online is quickest to start, but the local route is where the work happens, and the choice depends on whether you have a GPU.
Where Can I Download Baidu Unlimited OCR
Everything is open under the MIT licence, so you can use it commercially.
- The unlimited-ocr repository on GitHub is published by Baidu and holds the code, config and technical report
- The Baidu OCR Hugging Face page hosts both generations of weights, baidu/Unlimited-OCR and baidu/Unlimited-OCR-2, about 6.7 GB in BF16
- ModelScope mirrors both, much faster from mainland China
- Baidu Cloud and the Baidu AI open platform host the same model as a service
It supports vLLM and SGLang for serving, ms-swift for fine-tuning, and is now native in Hugging Face Transformers. The paper is on arXiv.
How to Use Baidu Unlimited OCR Free On HugginFace
You can see what Unlimited OCR 2 online free does without installing anything, without a GPU and without a Baidu account. The project README points to a browser demo on Hugging Face, and since the model has no hosted inference provider, that demo is the free online route.
-
Open the demo. Follow the link in the project README to the Hugging Face Space. There is nothing to install and no account to create.
-
Drop in a page. A photo, a scan or a PDF. Start with a single page, then click to parse document.
-
When the status is done, you can see the parsed text. Read the result. You can also get Markdown with headings, tables and formulas left in place, and the model marks where each block was found, so you can check it against the original.
-
Know the catch. The demo runs on a shared free GPU, so long documents queue behind other people, and your file lands on someone else's server. Treat it as a test bench, not a production tool.
How to Use Baidu Unlimited OCR Locally
If you tested and want to deploy Baidu Unlimited OCR from Github to your local device, you can try to follow the steps guide below:
-
Download the source code and model files from the official GitHub repository.
-
Set up the Python environment and install the required dependencies, including PyTorch, Transformers, and CUDA-compatible components.
-
Run the inference script to process images or PDF documents and save the OCR results locally. You can take scripts from guiding pages like guides and prompts from Marktechpost.
-
Deploy the model as a local API service using vLLM or SGLang if you want to integrate OCR capabilities into your own applications.
Most people mean the Baidu OCR text recognition endpoints when they say free OCR. The full list sits in the Baidu Cloud free tier docs.
Part 5. PaddleOCR vs. Unlimited OCR: What's the Difference?
This is the comparison most people want, so here it is in one table, and we have included PaddleOCR-VL because it is the middle option a two-way comparison hides.
Side-by-Side Feature Comparison
Please swipe to view
The pattern is clearer than the numbers. PaddleOCR is small enough to run in a browser or on a phone. PaddleOCR-VL is the precision option for a single dense page. Unlimited OCR is the reach option, The trade is a heavier model, a real GPU, and a habit of filling gaps it should leave empty.
Where Unlimited OCR Wins
- Page count, dozens of pages in one request and hundreds in version 2
- Cross-page structure, so a table split by a page break returns as one table
- Output that is immediately useful, clean Markdown rather than a bag of coordinates
- Licence terms, MIT with no negotiation needed for commercial use
- No per-page bill, because the weights are downloaded once and the hundredth page costs nothing
Where Unlimited OCR Struggles
- Single-page accuracy, where it sits below PaddleOCR-VL-1.6
- Hardware, since 6.7 GB of weights plus a GPU makes it a poor fit for laptops and phones
- Heading hierarchy, its weakest benchmark score, which the team itself flags as a design limitation
- Hallucination risk, because unreadable characters are completed rather than left blank
Part 6. Is Baidu OCR Safe to Use?
It depends on which of the three you use, and people miss that the cloud API and the open source model have very different privacy profiles.
Where Does Your Image Actually Go
Baidu Cloud OCR uploads your image over an encrypted connection, processes it on Baidu's servers, and stores parsed output in Baidu object storage with download links valid for 30 days. Baidu's privacy policy commits to collecting only what's needed, keeping data in China, and deleting it after the retention period, but doesn't publicly promise your data is never used for training---so read the agreement. Ignore community claims of no storage or 72-hour deletion; rely on the published policy's service-based retention period.
Options for Private or Sensitive Documents
Ranked by how much leaves your machine, there are four routes.
- Run the open source model locally, and the image never moves
- Offline SDK, where recognition happens on the device, for locked-down teams
- Private deployment, on your own servers, from around 75,000 yuan a year
- Cloud API, the simplest to run and the only one that uploads anything
For identity documents, medical records or anything under a compliance regime, we would start at the top of that list.
Part 6. PDNob: Smoother PDF OCR Workflow
If your actual work flow is to batch edit PDF or Images and OCR is just part of your workflow, PDNob can be a better choice for you.
PDNob PDF Editor is a desktop app that OCRs scanned pages and images, then lets you edit, convert, and annotate in one place. It turns photographed invoices or flattened contracts into selectable text, so you can reword lines, export tables to Excel, or save as Markdown---tasks most OCR tools leave elsewhere.
The OCR workflow takes four steps, and the official image and PDF OCR guide walks through all of them.
-
Launch PDNob PDF Editor, click "OCR PDF," select and open your PDF, and download the OCR module once if prompted.
-
In the OCR settings window, choose "Convert to Editable Text" for directly editable content or "Convert to Searchable Text in Image" to keep the original look while making text searchable. In Advanced Settings, enable only the cleanup options your scan needs---such as Auto Crop, Auto Deskew, Contrast Enhancement, Dark Spot/Noise Removal, and Text-on-Pictures Detection.
-
If you want, you can also upload images directly to PDNob and scan to copy text directly from images.
-
Click "Perform OCR" to run recognition, after which text becomes searchable and copyable, and in Editable Text mode you can correct individual words.
Conclusion: Picking the Right Product Line
Three Baidu OCR product lines solve different problems, and mismatching tool to job makes OCR feel unreliable: use the cloud API for production endpoints and structured fields, PaddleOCR for speed, offline work, or a small shippable model, and Unlimited OCR for long documents you need as structured Markdown. We've used all three, but whichever you pick, test it on your own documents first---yours are the only ones that count.
FAQ on Baidu OCR and Unlimited OCR
Q1. Is it safe for confidential documents?
A1: Only in the open source or privately deployed form. The cloud API uploads your files to Baidu, keeps parsed output for thirty days, and makes no promise about training use. If a document cannot leave your network, avoid the hosted endpoints.
Q2. Is Unlimited OCR free for commercial use?
A2: Yes. Both generations ship under the MIT licence, which permits commercial use, modification and redistribution. PaddleOCR is Apache 2.0, also commercial-friendly. Read the licence field on the model card you download. Read the licence field on the model card you download. The hosted API is a separate, metered service.
Q3. Unlimited OCR vs Unlimited OCR 2
A3: Version 1 shipped on June 22, 2026 and handles up to roughly 40 pages within a 32K context window. Version 2 shipped on October 27, 2026 and removes the practical page ceiling, with a 131K context window and about 6.7 GB of weights. Take version 2.
Q4. How many pages can one pass handle?
A4: There is no fixed published maximum. Baidu's demonstration parsed 3,623 pages in one request, and a separate test ran to 5,026 pages. In practice the ceiling is your context window and your patience: 131,072 tokens for version 2, plus a per-page cap in the serving config.
Q5. Does it support rare Chinese characters?
A5: Yes, on the high-accuracy tiers. The high-accuracy edition of general text recognition uses a character set of 20,000+ entries covering common and rare characters, and the edition with positions supports rare characters across 25 languages. The standard edition is a different product.
Q6. Can I run it on a consumer GPU?
A6: The model fits, but the memory story is what matters. Baidu's reported peak memory for one document fell from 47.00 GB to 1.77 GB in version 2, and the weights are about 6.7 GB. PaddleOCR is the easier answer on a laptop or phone: its tiny tier is 1.5 MB and runs in the browser at about 97 milliseconds per image.
- Make scanned PDFs searchable and editable with 99% OCR precision
- Batch convert PDFs to Word, Excel, PPT, images, PDF/A, Text, EPUB, etc., up to 30% faster
- Edit PDFs easily like Word, including text, images, watermarks, links, and backgrounds
- Annotate PDF with highlights, comments, shapes, stickers, and stamps
- Run smoothly on any PC without lags or crashes, even on low-spec machines
Secure Download
Secure Download
Speak Your Mind
then write your review
Speak Your Mind
Leave a Comment
Create your review for PDNob articles
Related articles
AI Tools Writer / Workflow Analyst
Rachel has spent the last 5 years testing and writing about AI-powered tools across writing, design, and research workflows. She focuses on practical AI tool usage, helping knowledge workers turn generative AI into everyday productivity.
