What launched, and when
Baidu released Unlimited OCR 2 on October 27, 2026. It is a vision-language model designed to convert document pages into text, including when several pages are submitted together so the model can process them with context across pages. The model card and project repository describe the release and provide usage instructions, but both come from the same team: they are primary documentation, not two independent confirmations.
The name “Unlimited” refers to the proposal to process documents over a very long horizon. It does not show that the system accepts documents of literally infinite length, or that every cost, service memory use, or total processing time stays constant as page count increases. The mechanism is described as a way to limit the set of attention references; practical scope also depends on configuration, available resources, and execution.
How it processes pages and documents
The released version supports both single-page and multi-page input. For a single page, the documented default inference path uses dynamic image tiling. In the documented multi-page path, each page is presented as a global view, and images are submitted in document order. The output includes one page marker for each input, in the same order. This organization may help preserve references across pages, but it does not prove that every layout element, table, or relationship between sections will be reconstructed correctly.
Baidu presents SR-SWA—Sparse Reference Sliding Window Attention—as the central mechanism in version 2. Its description says the method keeps a bounded set of working references for text-vision interactions. The aim is to contain attention growth in very long sequences. That explanation describes the proposed design; by itself, it does not establish that the total latency of a long job stays unchanged, because page reading, image conversion, output generation, and infrastructure also use resources.
Two documented input paths
| Use case | Input preparation | Relevant detail |
|---|---|---|
| One page | Page image | The described inference path uses dynamic tiling by default. |
| Multiple pages | Images in page order | The documented path processes them in one request and uses one global view per page. |
| Convert the PDF into images | The project’s example converts each page before submitting the images to the model. |
Access, execution, and cost
The documentation lists availability on Hugging Face, ModelScope, and Baidu Cloud, and describes self-hosted execution with Transformers or SGLang. This distinguishes downloading or running the model yourself from using a hosted service: a model’s presence on a platform does not, by itself, confirm terms of use, price, region, or commercial availability for a particular endpoint. The consulted information does not specify an official price for an Unlimited OCR 2 API.
The technical guide shows a test setup using Python 3.10 and an NVIDIA A800 with 80 GB of memory, along with dependencies and adapted SGLang components. This is a published example, not a universal minimum requirement. The serving configuration includes explicit limits for concurrency, tokens, and window size; anyone adopting it needs to account for their hardware, document length, and simultaneous workload. The guide also says that reducing the window can lower peak memory, at the cost of losing context when merging blocks that continue across a window boundary.
A practical workflow for a controlled test
- 01Set aside a representative document set and retain the original page images.
- 02If starting with PDFs, convert each page to an image while preserving the original order.
- 03Choose the single-page or multi-page path and record the crop mode and token limits used.
- 04Compare output against the source, paying particular attention to numbers, tables, headings, and continuity across pages.
- 05Measure resource use and errors in the same environment before increasing concurrency or window size.
What the published results show
The official material reports overall scores of 84.6 out of 100 on MPDocBench and 77.1 out of 100 on MarathonDocBench. These are figures reported by the developer for Unlimited OCR 2; the available evidence does not include an independent reproduction of those measurements and does not justify treating them as a guarantee for other documents, languages, formats, or configurations. A benchmark score summarizes its own dataset and evaluation rules, not the quality of every page processed.
The documentation also presents the version as a leader in long-document tests. That characterization should be understood as the provider’s claim unless an independently verifiable comparison under the same conditions and versions is available. The project cites an arXiv paper, but the available reading material does not include the paper’s contents or establish its peer-review status. It would therefore be inappropriate to attribute further conclusions to the paper or to say that the results have been replicated.
Limits to keep in mind
The serving example sets a global output-token limit and includes a per-page cap; the documentation notes that these and other parameters affect operation. So “can parse very long documents” does not mean every input will produce a complete transcription. An insufficient output budget, low-quality images, dense pages, or cuts between windows can affect the result. Output should be checked against the source, especially when it will be used for decisions or to extract sensitive data.
A bounded-attention mechanism should not be confused with a guarantee of constant cost for the whole system. Page count can still increase preprocessing and generation work, while service capacity depends on memory, concurrency, and parameters. Availability on the listed platforms does not settle commercial terms or whether a specific public API exists. The release therefore confirms publication and documented execution paths, while general accuracy, production costs, and independent performance require separate evidence.
What is documented—and what is not established
| Aspect | Documented | Not established |
|---|---|---|
| Release | Publication announced on October 27, 2026. | Independent validation of capabilities. |
| Long-document processing | Baidu describes SR-SWA and a multi-page path. | Infinite-length documents or constant total latency. |
| Benchmarks | The provider reports overall scores. | Guaranteed performance on other data or independent replication. |
| Access | Platforms and execution methods are listed. | A specific API price or the commercial terms of each service. |
| Hardware | An example uses an A800 with 80 GB of memory. | That this card is a universal minimum or that any system is sufficient. |
Open questions
- The consulted documentation does not specify an official API price for Unlimited OCR 2.
- The benchmark scores have not been independently reproduced in the available evidence.
- The available material does not establish the peer-review status of the arXiv paper cited by the project.
- Availability on listed platforms does not, by itself, establish commercial terms, regions, or the existence of a specific endpoint.
- The A800 with 80 GB of memory appears in a test configuration; it is not confirmed as a minimum requirement.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction