Unlimited OCR 2

Not yet independently verified. Rests on Baidu's own model card and repository, which date the release 27 October 2026. The state-of-the-art claims on MPDocBench and MarathonDocBench are Baidu's own, MarathonDocBench is a benchmark Baidu released the same day, and no independent dated test was found. We will update this when it can be confirmed, and remove this note.

Unlimited OCR 2 is a vision-language model from Baidu that turns page images into text, built for documents that run to dozens or hundreds of pages. Baidu released it on 27 October 2026, four months after the first Unlimited OCR, with open weights on Hugging Face under the MIT licence and a hosted version on Baidu Cloud.

What it does

Most OCR models handle a document one page at a time, which loses the context that runs across pages: a table that continues overleaf, a heading that governs several pages, a reference back to an earlier section. Unlimited OCR 2 takes all the pages of a document in one request, in order, and returns one parsed block per page.

The change in version 2 is an attention scheme Baidu calls Sparse Reference Sliding Window Attention. It compresses the interactions between text and page images from a very long sequence into a fixed set of working references, 512 tokens in Baidu's example, so that decoding time and memory per step stay nearly constant however long the document is. For very long files the repository adds page-window scheduling, which reuses the tail of the previous window. Baidu claims state-of-the-art results on MPDocBench, a benchmark of documents dozens of pages long, and on MarathonDocBench, a hundred-page benchmark it released the same day.

Who it is for

The model suits teams that extract text and structure from long contracts, filings, manuals or scanned archives and want to run the extraction themselves rather than send documents to a hosted service. The card gives inference through Transformers for testing and SGLang for serving, tested on a single 80 GB data-centre GPU, and the ModelScope ms-swift toolkit supports fine-tuning it.

Key facts

Maker: Baidu Released: 27 October 2026 Parameters: about 3.3 billion (per the Hugging Face weights listing) Licence: MIT Inputs: page images; output is text, one block per page Available on: Hugging Face, ModelScope, Baidu Cloud Price: open weights; Baidu Cloud pricing not read