Baidu Opens Unlimited-OCR-2, a 3B Model for 100-Page Documents

Baidu released Unlimited-OCR-2 on Oct. 27, an MIT-licensed document-parsing model that its ERNIE team says reads a hundred-page file in one pass, with every benchmark figure so far supplied by Baidu.

The model card on Hugging Face lists 3 billion parameters and asks users to run custom code with trust_remote_code. It shows about 189,897 downloads in the last month, no gating, and no inference provider hosting it. The card does not state the training data.

What SR-SWA changes

The release swaps the decoder attention of the first Unlimited OCR for Sparse Reference Sliding Window Attention, or SR-SWA. The card says it limits attention to "a fixed-size set of working references," for example 512 tokens, so the cost of each decoding step stays roughly flat as a document grows. The paper, arXiv:2610.39308, says that at the hundred-page scale each step is 7.0 times faster than full attention and 5.6 times faster than the first model's R-SWA.

The first model, covered by The Decoder on July 5, kept only the last 128 output tokens in view and was built on the open DeepSeek OCR. The Decoder quoted its authors: "No current OCR model handles more than about ten pages in a single pass." That article gave the first model about 3 billion total and 500 million active parameters. A Let's Data Science report says Baidu's release material does not say whether version 2 keeps that size, so the 3 billion on the card is the only figure for the new model.

What the benchmarks do and do not show

Baidu says the model is state of the art on two tests. Let's Data Science reports the scores below, and says Baidu ran both and released MarathonDocBench alongside the model. The card carries the leaderboard only as an image, with no numbers in text.

Benchmark	Score	Who ran it
MPDocBench, multi-page parsing	84.6	Baidu
MarathonDocBench, about 100 pages	77.1	Baidu, on its own new test

Let's Data Science cites Pandaily as reporting that independent evaluations had not been published at the time of writing. The Terminal found none either.

1. SR-SWA vs full attention7 x
2. SR-SWA vs R-SWA5.6 x

Source: Unlimited OCR Works 2, arXiv:2610.39308, accessed 2026-11-03

Limits and a date conflict

The Decoder's report on the first model found that its fixed 32,000-token context caps how many pages fit, since visual tokens accumulate with each page, and its headline said the model was not truly unlimited yet. Whether version 2 lifts that cap is unclear from the card.

Let's Data Science says Baidu's README tunes serving for 10-page scheduling windows. Smaller windows lower peak memory but weaken merging across pages, and a memory fraction of 0.7 was stable at 32 or more concurrent requests where 0.75 was not. The card's Transformers tests ran on one NVIDIA A800 with 80 GB.

The sources disagree on dates. The card dates the release 2026/10/27 and the paper 2026/11/02, while arXiv lists the paper as submitted Oct. 29 and revised Oct. 30. Two other open-weight releases covered here, Qwen3.8 27B and Mistral Large 4, are general language models rather than document parsers. The Decoder said Baidu plans 128,000-token models after the first release, and gave no date.