What happened
Baidu AI Cloud’s OCR documentation, updated on October 30, 2026, adds a hosted capability called Document Parsing (Unlimited-OCR). The supplied documentation describes it as an API built on Unlimited OCR 2, Baidu’s document-parsing model. Rather than presenting the model only as software a team must run itself, Baidu is exposing it through its cloud OCR product surface.
The documented task is document parsing rather than plain text recognition alone. Baidu says the API can merge tables and paragraphs across page boundaries. That detail matters for multi-page PDFs: a table split by a page break, or prose continuing onto the next page, is a structural problem that page-by-page OCR pipelines must reconstruct after the fact. The API returns results in JSON or Markdown, two forms that can be passed into downstream extraction, retrieval, review, or automation workflows.
The interface is asynchronous. A client submits a job and then polls for its result, rather than receiving the parsed document in the initial request. The documentation states a limit of 2 QPS for each endpoint. That is enough to make the service testable in an application integration, but it is also a concrete operational constraint for teams considering large backlogs or interactive user experiences. The supplied material does not specify latency, document-size limits, supported file types, regional availability, or output-quality guarantees.
Baidu is offering the API free for a limited time. Verified individual users receive 200 free test pages; verified companies receive 1,000. The source does not give an end date for the promotion or a subsequent price. The release is therefore an access change with a bounded trial allocation, not evidence of a permanent free tier. Builders can use the allocation to determine whether cross-page reconstruction and structured output work on their own forms, reports, contracts, or scanned archives.
What the archive adds
This is a new access event for a known story. Earlier coverage described Baidu open-sourcing Unlimited OCR 2 as a 3B-parameter, MIT-licensed model for documents of any length, and placed it in the lineage of Baidu’s first Unlimited OCR release and related OCR systems. The new fact is the hosted API, its asynchronous workflow, and its trial limits.
Trend signal
No supplied pattern genuinely fits this release. It is a product-access change for document parsing, not evidence on a broader agent architecture or operating-model thesis.
Connections to other events
The API is related in subject, but distinct in event type, from Introducing the Document Parsing Protocol, which described LlamaIndex’s draft shared answer key for parsing judgments. Baidu is exposing one hosted parsing capability; the protocol proposal concerns how parsing decisions might be represented or evaluated across systems.
Why it matters
Many AI workflows begin with documents that are not usable inputs: scanned PDFs, reports with multi-page tables, and records whose reading order must be reconstructed. A hosted API can lower the integration cost compared with operating an open model, especially for a prototype that needs structured JSON or Markdown quickly. The cross-page claim is the feature worth testing because it targets a common source of downstream errors.
The operational trade-off is clear even from the limited documentation. Teams gain a managed endpoint and structured output, but accept asynchronous polling, a 2-QPS endpoint limit, verification requirements for trial pages, and an unspecified future price. This should be treated as a candidate component in a document pipeline, not as a demonstrated replacement for validation. A team’s own documents, languages, layouts, and error tolerance determine whether it is useful.
Limits and uncertainty
The supplied source is product documentation, not an independent evaluation. It does not provide accuracy figures, benchmark comparisons, latency measurements, supported languages, maximum page counts, file-format coverage, data-handling terms, or a date and price after the limited free period. It also does not establish that cross-page tables and paragraphs are reliably reconstructed on every document class. The service may be useful for experimentation, but the evidence here is insufficient to judge production quality, scalability, geographic access, or total cost.
Skill to build now
Build a small evaluation set before integrating any document-parsing API: include a multi-page table, a paragraph spanning a page break, a scanned page, and the document type that drives your workflow. Submit the set, preserve the JSON and Markdown outputs, and compare them with a human-checked answer. Record structural failures separately from character-recognition errors. That makes the advertised cross-page capability measurable and reveals whether asynchronous polling and the 2-QPS limit fit the intended product flow.
Reader takeaway
Unlimited OCR 2 is now available not only as a model but through a Baidu AI Cloud document-parsing API. The practical opportunity is fast testing of structured, cross-page document extraction; the practical discipline is to validate it on real files before relying on its output or its temporary free allocation.
Open question
Will Baidu publish permanent pricing, supported-document limits, and evidence that cross-page structure is preserved reliably across common business documents?