The MIT license lets banks, insurers and law firms build Baidu's long-document reader into their own products and run it on their own servers without paying Baidu.
Baidu has open-sourced Unlimited OCR 2, a 3-billion-parameter model that turns scanned and photographed pages into text and can work through hundreds of pages in a single request. The model card and the GitHub repository date the release to October 27 and publish it under the MIT license, which allows commercial use. Sina Tech and Leiphone reported that Baidu showed the model reading the seven Harry Potter books, more than 3,600 pages, in one pass.
OCR, or optical character recognition, is the step that converts page images into machine-readable text, and it sits underneath most corporate uses of large language models: contracts, annual reports, invoices, insurance claims and research archives. Baidu’s earlier version, released in June with a 32,000-token context window, topped out at a little over 40 pages. The new release is built to remove that ceiling by keeping the model’s working memory at a fixed size, however long the document runs.
The headline results come from Baidu’s own tests on test sets Baidu designed, and no outside evaluation had been published by the time of the release. Chinese coverage brands the model under Wenxin (文心), the Chinese name of Baidu’s ERNIE family, although the English model card makes no reference to ERNIE.
A fixed-size working memory lets the model keep reading
Most language models slow down and need more memory with every page they read, because they keep a growing record of everything seen so far. Baidu’s answer is a technique it calls sparse reference sliding window attention, or SR-SWA. According to the model card, it squeezes the link between the page images and the text already produced into a small, fixed set of reference points, for example 512 tokens, so the time and memory needed for each new step stay roughly flat as the document grows.
In its default setup, the model reads 10 pages at a time and can write up to 20,000 tokens for each page. Sina Tech and Leiphone reported Baidu’s figures that, on hundred-page documents, each decoding step runs seven times faster than with standard attention and 5.6 times faster than the previous generation’s approach.
The memory savings are larger still. Xinzhiyuan, a Chinese AI media outlet, reported that peak working memory for a single long document fell from 47 GB to 1.77 GB, a cut of about 96%. Serving 16 requests at once, throughput rose from 360 to 1,315 tokens per second. For a company processing thousands of filings, those figures translate directly into fewer graphics processors and lower cloud bills, if they hold up outside Baidu’s lab.
Baidu reports a wide lead on very long documents
Baidu scored the model on two benchmarks. On MPDocBench, a multi-page parsing test, Unlimited OCR 2 scored 84.6. On MarathonDocBench, which uses documents of 103 to 300 pages, it scored 77.1, according to Sina Tech and Leiphone. Xinzhiyuan reported that the next-best system on MarathonDocBench scored 33.74 and that Google’s Gemini-3.1-Pro scored 26.97, a gap that reflects how quickly general-purpose models lose accuracy over very long inputs.
The first Unlimited OCR, released on June 22, drew on DeepSeek’s open OCR work and posted 93% on the public OmniDocBench v1.5 test, The Decoder reported in July. Baidu has not published OmniDocBench results for the new version, which makes a like-for-like comparison with other open models harder.
The project already has a large developer following. Sina Tech and Leiphone reported more than 10 million downloads on Hugging Face and about 28,500 stars on GitHub, where the repository showed close to 28,900 stars and 3,100 copies made by other developers when EastFrontier checked it. The repository credits DeepSeek-OCR, DeepSeek-OCR-2 and Baidu’s own PaddleOCR among its foundations, a reminder of how much Chinese open-source labs now build on one another’s releases.
Companies can test it on a single data-center GPU
The model card lists testing on one Nvidia A800 with 80 GB of memory, and the weights ship in a format that standard tools can load directly. Baidu supports running it through Hugging Face Transformers, the SGLang and vLLM serving engines, compressed versions for llama.cpp, Ollama and LM Studio, ModelScope and Baidu’s own cloud. That breadth of options lowers the cost of a trial for an in-house document team.
The open license matters most for banks, insurers, law firms and government agencies that cannot send sensitive files to an outside service. They can now run the model on their own hardware and fine-tune it for their own forms without paying Baidu a license fee.
Sina Tech and Leiphone reported that Baidu will show more of the technology at Baidu World on November 12. Before then, the clearest test will be whether independent researchers reproduce the long-document scores, or post results on public benchmarks such as OmniDocBench that let buyers compare the model with rivals.