Futureweb

Software

The OCR 2 benchmark is the repo

Baidu says Unlimited OCR 2 landed on October 27. The result line is the project claim, on a benchmark the project also released that day. It is not an outside table.

The short version

The repository's October 27 note claims the result, including on MarathonDocBench, which the same note says the project released. A project's table is the project's.

The release note

The baidu/Unlimited-OCR repository's release list says Unlimited OCR 2 was published on October 27, 2026. The Hugging Face card uses the same date and the model id baidu/Unlimited-OCR-2. The repository license is MIT. The README says the new version is meant to parse documents whose length is not capped by a growing attention cost, and it names a fixed-size set of working references as the reason the per-step cost should stay roughly flat. That sentence is the project's description of its own method. This page will not expand it into a design note, and it will not copy the install lines the README includes.

The benchmark the project shipped

The same October 27 lines announce MarathonDocBench, described as a hundred-page parsing benchmark, and they claim state of the art on that set and on a shorter set the project calls MPDocBench. A table the author publishes, on a test the author also publishes, is a result the author measured. It can be true and still not be a third-party league. Nanbeige's table was filed on this distinction for a different lab. The ffmpeg pin was a version choice, not a claim that the upstream project had been replaced. OCR 2 does not replace every other OCR system by appearing in a repository.The ffmpeg pin.

MarkTechPost is the wire account

MarkTechPost's item is how the release reached the wire. The article describes the repository. It also prints setup commands. Those commands are omitted here. A reader who wants to run the model has the repository and the model card. A reader of this brief needs the date, the license, the model id, and the status of the benchmark. The Node October release was a patch level with a vendor changelog. This is a research repository's changelog. Both are software documents. Only the Node release is a runtime a general user was already on.The Node October patch.

What the card adds

The model card repeats the October 27 release line and adds distribution notes: a ModelScope copy, a Baidu Cloud listing, and a demo. Availability on a cloud is the project's statement that a hosted copy exists. It is not a support contract, and it is not a measurement of accuracy on a customer's documents. The xml pin was a parser crate beside libxml2. A document OCR model is a parser of a different kind. The pin brief is the reminder that a new parser does not retire the old ones by announcement.The xml pin.

d3-ultra is a different card

d3-ultra, filed today on the artificial-intelligence desk, is a decision model whose test questions are private. Unlimited OCR 2 publishes a benchmark the project also wrote. Privacy of a test set and authorship of a test set are different limits, and both mean the number is not an independent ranking. The h264 pin is a codec choice. It is linked so a model release is not read as a codec bake-off either.The h264 pin.

The KVM pull is pre-release mail

Today's KVM brief is a maintainer pull aimed at Linux 7.4, which is not yet a tarball. Today's JetBrains brief is an IDE parser swap. OCR 2 is a tagged release in a public repository, dated in the README itself. Three software filings were required because the day had three documents and the other desks could not absorb them without merging unrelated dependencies. They stay three briefs. Asked directly, What does the project claim changed? The README says a fixed-size set of working references is meant to keep per-step cost from growing with the length of the document. That mechanism claim is the project's, and this page does not expand it.The KVM pull.

Who it is for

The repository addresses people who parse long documents and who can read a research release. It does not say a particular scanner, hospital, or bank has switched. The first Unlimited-OCR line in the same README is dated June 22, 2026, so version 2 is a second release, not the project's first appearance. A reader who wants the June model has a different date in the same file.

What this page will not claim

Flat cost as documents grow is the project's complexity claim. This page did not time a hundred-page file. State of the art is the project's score against the systems it chose to compare. A later independent table can move the claim. Until that table exists, the October 27 README is the record of what Baidu says, not of what a referee measured. Asked directly, What date does the repository put on Unlimited OCR 2? The release list in the repository says October 27, 2026. The model id on the card is baidu/Unlimited-OCR-2. The license on the repository is MIT.The JetBrains terminal.

The record

The record is the GitHub repository for the October 27 release note and the license, the Hugging Face card for the model id, and MarkTechPost for the wire account. The benchmark is the project's. The install steps stay in the repository. The fixed-size reference sentence stays the project's description, unexpanded here. The short answer this page will stand on is this: The repository's October 27 note claims the result, including on MarathonDocBench, which the same note says the project released. A project's table is the project's.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Repository
  2. Model card
  3. MarkTechPost

Questions

What date does the repository put on Unlimited OCR 2?

The release list in the repository says October 27, 2026. The model id on the card is baidu/Unlimited-OCR-2. The license on the repository is MIT.

What does the project claim changed?

The README says a fixed-size set of working references is meant to keep per-step cost from growing with the length of the document. That mechanism claim is the project's, and this page does not expand it.

Who published the hundred-page benchmark?

The project. The card's October 27 line says MarathonDocBench shipped with the model. A benchmark the author releases is not an independent league table.

Does MarkTechPost add a second measurement?

It reports the repository. It is not a lab that reran the pages. Where the article adds a how-to, this brief leaves the how-to out.

Is this the JetBrains terminal or the KVM pull?

No. Those are an IDE parser and a kernel merge. This is a document-parsing model the repository says it released.