---
title: "Baidu releases Unlimited-OCR 2 for long-document parsing"
description: "Baidu releases Unlimited-OCR 2 on October 27. Explore open weights, hosted API access, long-document parsing and practical deployment limitations."
url: "https://promptszone.com/baidu-unlimited-ocr-2-long-document-model-release"
published: "2026-10-28T00:00:05Z"
categories: ["New Models"]
---

# Baidu releases Unlimited-OCR 2 for long-document parsing

Baidu's October 27 model release targets cross-page document parsing with open weights and a hosted beta. Deployment trade-offs and service limits still matter.

![AI-generated conceptual illustration of pages flowing from an open book into organized text and table blocks; not actual OCR output.](https://promptszone.com/assets/uploads/495c7023d1e4125e1e749559be39ecc5.png)

**Summary:**

Unlimited-OCR 2 introduces sparse-reference attention for long documents. Its model release and hosted API offer different access paths, with important limits to verify.

Baidu released Unlimited-OCR 2 on October 27, pairing an open-weight document-parsing model with a hosted public beta. The update targets a practical weakness in document AI: keeping information connected across long files rather than treating every scanned page as an isolated image.

The [official project repository](https://github.com/baidu/Unlimited-OCR) dates the release to October 27 and links the model, a demonstration and Baidu Cloud access. This is a model release with deployment routes, not merely a promise of future weights.

## A different approach to long documents

Baidu calls its new mechanism Sparse Reference Sliding Window Attention, or SR-SWA. The team says it retains a bounded set of working references instead of letting attention costs expand unchecked with document length. Its claims concern steadier decoding costs; they do not mean an arbitrarily large document becomes free or instantaneous to process.

The project also introduces MarathonDocBench for very long documents. Baidu reports leading results on long-document benchmarks, but those are developer-reported evaluations, not independent evidence that the model will outperform every alternative on a reader's own files. PromptsZone has not run the model or reproduced its benchmarks.

## What developers can use now

The [official Hugging Face model card](https://huggingface.co/baidu/Unlimited-OCR-2) lists a roughly three-billion-parameter model under an MIT license, with Transformers instructions for experimentation and SGLang guidance for serving. Its example setup was tested on an NVIDIA A800 with 80GB of memory; that is an example configuration, not a verified minimum hardware requirement.

Deployment still involves trade-offs. The serving notes warn that smaller processing windows can reduce peak memory while weakening the merging of material across window boundaries. Teams should evaluate page continuity, not just whether individual words look correct.

Raw model output is not automatically a polished Markdown document. The project's [post-processing documentation](https://github.com/baidu/Unlimited-OCR/blob/main/utils/README.md) describes layout blocks and a conversion utility that reconstructs headings, groups continuing text and merges table fragments. That distinction matters when connecting OCR output to search, retrieval or downstream automation: the conversion stage is part of the system being evaluated.

## “Unlimited” does not remove service limits

Baidu's [hosted API documentation](https://ai.baidu.com/ai-doc/OCR/fmr1p39gb), updated October 27, describes asynchronous document parsing with JSON or Markdown results and cross-page table and paragraph merging. It separately limits PDF submissions to 500 pages. The model's long-horizon positioning should therefore not be read as an unlimited hosted-service allowance.

The documentation advertises a limited-time free test allocation tied to verified accounts: 200 pages for individuals and 1,000 for enterprises. These are conditional testing allowances, not a permanent free tier or a guarantee that every reader can immediately access the service. Eligibility and current terms should be checked before building around them.

For organizations digitizing archives, the useful question is whether the complete pipeline preserves reading order, headings and continued tables reliably enough for the intended workflow. Unlimited-OCR 2 offers a new option to test, with downloadable weights and hosted access serving different deployment needs. Neither route removes the need to inspect errors and validate important extracted information against the original document.

Image: AI-generated conceptual illustration created for PromptsZone; not Baidu product imagery or Unlimited-OCR 2 output.