The limits of Unlimited OCR 2

Baidu’s October 27 release puts a fixed-reference attention design beside explicit page windows and output caps. The documents explain the tradeoff; independent performance remains unverified.

Reporting · 2026-10-30 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · Baidu’s October 27 release record introduces Unlimited OCR 2 as a model for parsing very long documents. Its public model page lists three billion parameters and an MIT license, while the file tree exposes two weight shards within a 6.87 GB repository. Baidu also provides code and inference instructions. The release gives readers something more concrete than a demonstration: a named model, downloadable artifacts and a documented way to submit multiple pages. Those records establish what Baidu has made available; they do not measure its reliability in another organization’s workflow. [1] [2] [3]

What stays in working memory

Primary-source statement · The project’s June technical report describes the problem behind that design. In the authors’ account, a decoder’s accumulated key-value cache consumes more memory and slows generation as output grows. Their earlier Unlimited OCR model replaced its decoder attention with Reference Sliding Window Attention. The October release describes a sparse version, SR-SWA, which compresses interactions between visual material and text into a fixed set of working references. Baidu gives 512 tokens as an example and claims nearly constant memory and latency for each decoding step. [4] [1]

Primary-source statement · Baidu’s multi-page example turns each page into an image, places the images in order and submits the document in one request. The model card says the resulting text contains a page marker for each input page, preserving that order. Its serving instructions also describe page-window scheduling for very long documents, with the previous window’s key-value tail reused. These are documented behaviors and interface examples. They do not show that every table, paragraph or page boundary will be reconstructed correctly in an unfamiliar document. [3] [1]

Primary-source statement · The documented software environment is specific. Baidu lists Python 3.10, PyTorch 2.13.0 and Transformers 4.52.1, tested on an NVIDIA A800 with 80 GB of memory. The two published weight shards are listed at 5 GB and 1.69 GB; the larger repository total includes other files. These package sizes are different from the memory needed to run a workload. The instructions use custom model code and a customized SGLang wheel. Neither stack was installed or executed for this report, and the listed hardware is not a verified universal minimum. [1] [2]

The limits inside the release

Primary-source statement · The serving example makes the operational choices visible. Its page windows are set to ten pages, with a per-page output cap of 20,000 tokens. Other settings control concurrency, memory reservation and temporary allocations. Baidu warns that a smaller page window lowers peak memory but weakens merging across pages because a continuing block loses context at the boundary. The release’s unlimited-horizon description therefore sits beside explicit resource settings and a documented context tradeoff. It is not evidence that every document can be processed without any limit. [1]

Primary-source statement · The paper linked from the model page predates this release. Its arXiv record was submitted on June 22, and its abstract describes the earlier R-SWA design and transcribing dozens of pages within a 32,000-token maximum. That record supplies the project’s architectural lineage. It does not independently verify the October model, establish version two’s release date or demonstrate that its newer sparse attention has the same behavior in every setting. The dated October entries in Baidu’s repositories are the relevant release records for this edition. [4] [1]

Primary-source statement · Baidu claims leading results on both dozens-of-pages and hundred-page evaluations, naming MPDocBench and MarathonDocBench. Those are issuer performance claims, not results reproduced by ICU. The reviewed documents do not establish independent error rates across different layouts, languages, handwriting or production loads. The repository’s warnings about cross-page merging are particularly relevant to the promised whole-document use: the company itself identifies a setting that changes how much context survives. No benchmark score, throughput or cost in this article should be read as an independently measured result. [1] [3]

Primary-source statement · The public package lets a reader inspect a narrower proposition than the model’s name alone conveys. Baidu describes a fixed-reference attention design, publishes weights and code, and exposes choices about page windows and output budgets. What remains unverified is the resulting accuracy, speed and reliability across independently tested documents. The useful question for the next evidence is precise: when working memory is bounded and a document crosses those windows, how much of its structure and context does the system retain? [3] [1]

Reporting note. Baidu originals and the project’s June arXiv record were checked October 28. All four records belong to one project origin. Technical and performance statements are attributed; no installation, model execution, security audit, paid call or independent benchmark was performed.

Corrections

No corrections recorded.

Report an error