*Conceptual illustration of the scenario; not a product interface, client project, or test result.*
Baidu has introduced Unlimited OCR 2, an open model for parsing long documents while preserving cross-page relationships. The official announcement is dated October 27, 2026. For business readers, the interesting part is not the slogan about “infinite pages,” but the ability to handle a table or paragraph that continues after a page break. At the same time, the technical report openly acknowledges a limitation: the mechanism does not preserve all global document information, which makes it harder to understand the overall structure. Baidu announcement, technical report.
An overview for document workflow owners, enterprise search teams, and IT departments with large archives. Source materials were checked on October 28, 2026. We did not run the model, recognize our own PDFs, or measure Russian-language accuracy, speed, or GPU memory usage. Developer claims and editorial scenarios are separated below; there are no ready-to-use installation commands here that have been verified by our editors.
The main news: a way has been proposed to reduce the computational load of one step when parsing a long document. The main caveat: this does not mean the processing time is the same for a short file and a long one, or that the hierarchy of sections will always be correct.
Contents
- What Baidu released
- Why cross-page links matter
- What Unlimited means
- What the tests prove
- Local deployment
- Role in RAG
What Baidu released exactly
In the repository and model card, the second version release is marked as October 27. Available assets include baidu/Unlimited-OCR-2, code, and materials about the method. The source repository is published under MIT, and the same license is listed in the model card metadata. This opens the door to independent study and deployment in compliance with the license terms. Model card, license text.
OCR is text recognition from an image; document parsing is broader because it also includes reading order and content structure. In its announcement, Baidu describes recognition of page elements and merging of continuing tables and paragraphs. These features should be understood as the developer’s product claims, not as editor-verified performance on any and all scans. Baidu feature description.
This news stands out especially against the familiar page-by-page pipeline. If each sheet is processed independently, the continuation link has to be reconstructed by the next stage. In a multi-page report, that can be just as important a task as reading individual characters.
This is a new parser, not a ready-made enterprise document management system. Archive access rights, source file storage, search, error correction, and data loading into a DMS or ERP remain part of the surrounding process. The open model lets you replace one component of that process.
Why cross-page links matter for business
Imagine an editorial example: a specification ends in the middle of a long table. On the next page, the rows continue, but the object name and column heading stayed on the previous page. If you read the pages separately, you can preserve every number and still link them to the wrong product or unit of measure.
Another example is an instruction paragraph in which the condition stayed at the bottom of the page and the action moved further on. For search inside such a document, it is important not only to extract the text, but also to preserve the connected sentence. Otherwise, the system will find the right words, but the excerpt will not be enough to answer the employee.
| Archive situation | What kind of linkage is needed | What remains a separate task |
|---|---|---|
| A table continues on the next sheet | Preserve the columns and row continuation | Check values and units of measure |
| A paragraph is split by a page break | Restore continuous text | Identify the semantic boundary of the section |
| The report contains many subsections | Match the text to its position in the document | Build the correct heading hierarchy |
| Documents are prepared by different departments | Obtain a comparable structured result | Normalize names and access rules |
All table rows are analytical examples, not results of the Unlimited OCR 2 test. They show why comparing parsers on a single clean page may not answer the needs of a large archive. If short, repetitive forms dominate in your company, the main effect of the new multi-page mechanism may be less significant.
So the reason to pay attention is the real content breaks in your document class. The page count by itself still does not explain why you should switch tools. It matters more to understand whether the system loses row links, paragraph continuity, or the source of an answer.
What Unlimited means and how SR-SWA works
SR-SWA stands for Sparse Reference Sliding Window Attention — an attention mechanism with a limited set of active reference data. According to the authors, it reduces the amount of visual information the model consults at a single text generation step. As a result, the cost of that step should not grow the way it would when accessing the entire accumulated visual context. Method description in the repository.
This is the difference between the cost of a single step and working across the entire document. A long file contains more pages, images, and output text. Even if each step costs roughly the same, there will be more steps. Therefore, the architectural claim does not imply a fixed processing time for any PDF.
The second difference is local coherence versus global understanding. In the limitations section of the report, the authors say the current mechanism does not preserve global document information and limits understanding of structures such as heading hierarchy. That is an important caveat alongside the promise of handling cross-page continuations. Study limitations.
For business users, the takeaway is specific: getting the continuation of a table and correctly identifying which section of the report it belongs to are different outcomes. If the end goal requires a section tree, you cannot assume it is automatically solved by switching to Unlimited OCR 2.
What the benchmarks show and what they do not prove
The authors evaluate the model on MPDocBench, where documents reach 30 pages, and on their presented MarathonDocBench with documents from 100 to 300 pages. The evaluation includes several metrics: text, reading order, tables, formulas, images, and cross-page links. The final score aggregates these different quality dimensions. Report methodology and results.
Therefore, a high final score cannot simply be renamed as a “percentage of correctly recognized fields.” It is a different unit of evaluation. The report was prepared by the developer’s team; this article does not include an independent reproduction of the results.
Accuracy on your Russian archive is not established either. Even if a specific model is labeled multilingual, that is not enough to conclude it can read small footnotes, handwritten corrections, stamps, or damaged scans. For different workflows, errors will matter differently: in one case the search snippet is affected, in another the value in an accounting system.
The editorial takeaway on benchmarks is that they explain why the model is worth evaluating, but they do not determine whether it is ready for your workflow. For an archive of specifications, what matters are table continuations and the connection to the source page. For a book, it is the reading order. For a management report, the structure of sections may be critical, and the authors acknowledge that limitation.
Can Unlimited OCR 2 be run locally?
The official materials show two paths: Transformers for quick evaluation and SGLang for serving requests. The Transformers example uses CUDA and BF16; the developers specify a test environment with an NVIDIA A800-80GB. For SGLang, a dedicated build from the project is offered. These details describe the published environment, but they do not establish the minimum VRAM for any document. Developer instructions.
In the multi-page example, an ordered list of page images is passed in. In the server example, the PDF is first converted into images. So the wording “parse a PDF with one request” does not mean you can skip page preparation and send any arbitrary file to any standard interface. Document processing examples.
For an IT team, there are two separate assessments here: whether the published code can be run, and whether a production queue can be served reliably. A large document takes longer to process than a short one, and concurrent requests change resource usage. The model name does not define the server’s available throughput.
Local deployment is useful as a hosting option, but it is not a ready-made data protection scheme. You will need a chosen environment, access control, result storage, and rules for updating dependencies. We do not provide our own installation command, do not promise compatibility with a specific workstation, and do not present Baidu’s published examples as a run performed by our editorial team.
How the release affects RAG and enterprise search
RAG is the retrieval of relevant passages from a knowledge base before the model generates an answer. In this process, the parser creates the source material for the index. If a table is split incorrectly or paragraph linkage is lost, the answer generation step is already working from a distorted foundation.
Unlimited OCR 2 is interesting as a candidate at the extraction stage, especially for long scans with continuations. But the recognized document still needs to be split into meaningful chunks, linked to its source and version, access rights have to be taken into account, and search must be prepared. The analytical conclusion from the global-structure limitation is this: when building sections, you cannot rely only on the resulting text stream without verification.
For a team that already has a parser, it is more useful to compare the nature of the losses than to change the entire architecture just because of the release name. For example, the current tool may read individual pages well but fail to merge tables; another issue may be incorrect section detection. These are different reasons to change the pipeline and different requirements for the next component.
The broader context for the tools is covered in our comparison of PDF parsers for RAG. Preparation of paper sources is discussed separately in our archive digitization checklist. These tasks remain relevant even after a new OCR release.
How AIrassvet connects long documents to a knowledge base
Working with long documents can include preparing page images, integrating recognition with a corporate repository, and building RAG search with fragment-to-source linking. For multi-page tables, the core task is to preserve row and page relationships, and for search, to provide employees with a verifiable source for the answer. The choice of a specific parser and its deployment conditions are determined separately.
It is best to start with one document type: describe its structure, the current losses during processing, storage constraints, and the expected result. Discuss the project.
What Unlimited OCR 2 means for business
Baidu’s release advances an important direction: long documents can be processed while preserving continuity of content across pages, reducing the computational burden of any single step. For business, this is a reason to evaluate the new extraction component if your current process loses those links. The word “unlimited” in the name does not eliminate processing time, infrastructure limits, or the authors’ acknowledged global-structure issue. Results on a Russian archive remain a matter for separate verification.
