An entire book read in one pass
China's Baidu on Oct. 27 released on Hugging Face the weights of "Unlimited-OCR-2," an AI model that reads text and tables from document images. It is released under the MIT license, which places few restrictions on use, and can be used commercially. @aitrackerbot, an X account that tracks new model releases, was among the first to report the release.
The small model has about 3 billion parameters and, according to Baidu, can read documents of tens to hundreds of pages in a single pass without splitting them into individual pages. Baidu also showed an example of processing all seven Harry Potter books, more than 3,600 pages, in one pass.
Speed holds up as documents get longer
AI models that read documents usually slow down as pages are added, because the computation needed to check against earlier content grows. Unlimited-OCR-2 uses a mechanism called "SR-SWA" that narrows the portion it references to a fixed number of about 512. A gate-like component that selects pages lets the model look only at relevant pages. Baidu says this keeps the time and memory needed to output each character nearly constant regardless of document length.
Under a 100-page setting, it is 7 times faster per step than an approach that references everything, and 5.6 times faster than the approach of the first version released in June, according to Baidu. Baidu said the original Unlimited-OCR had been downloaded more than 10 million times on Hugging Face.
| Benchmark | Description | Score |
|---|---|---|
| MPDocBench | Reading documents of tens of pages | 84.6 |
| MarathonDocBench | Reading documents of about 100 pages | 77.1 |
Evaluations are Baidu's own
Baidu says the model ranks first on both benchmarks among approaches that read a document from start to finish with a single model. However, it did not give the names or scores of the models it compared against. MarathonDocBench is a benchmark Baidu released alongside the model, and no third-party evaluation has been published yet.
Even with constant speed, memory still needs tuning. The model's documentation cautions users to allocate memory conservatively when processing 32 or more requests at the same time. Baidu plans to present the model in more detail at its "Baidu World 2026" event on Nov. 12.
