Baidu has open-sourced Unlimited-OCR-2, a document-parsing model designed to handle very long files without requiring the entire document to fit inside a conventional context window. Its approach could make multi-page analysis more practical, while support for Transformers and SGLang gives developers familiar routes to try it.
Watch Desk analysis
What happened
Baidu’s model page describes Unlimited-OCR-2 as using Sparse Reference Sliding Window Attention (SR-SWA) to manage computation and memory while parsing long documents. It supports cross-page context and dynamic tiling for individual pages. The model page also says the release includes native support for Transformers and SGLang.
Baidu presents the model as capable of “unbounded” document parsing within a finite context window and reports strong results on extreme-length benchmarks, including MarathonDocBench and tasks involving documents of hundreds of pages. Those are the developer’s claims; the supplied information does not include benchmark results or independent evaluations.
Why it matters
Long reports, research papers and document collections can exceed the practical limits of a model’s context window. A system designed to carry information across pages while managing memory could help developers build tools that search or analyse larger documents without simply trying to stuff everything into one prompt. That is a useful technical problem to tackle, even if “unbounded” is doing some ambitious work in the description.
The stated support for Transformers and SGLang matters too: developers can explore the release through existing model and serving ecosystems, rather than treating it as a paper-only idea. How well it performs on real documents, and what hardware or setup it needs, are not established by the supplied details.
Our read
This is a meaningful open-model release because it targets a recognisable bottleneck and names the mechanism and serving options, not just a grand claim about reading everything. Baidu’s benchmark claims are a starting point, not a verdict. The practical test is whether the model can preserve the details that matter across long, messy documents without turning memory management into a new puzzle for its users.
What to watch
- Independent results on MarathonDocBench and other long-document tasks.
- Whether the model’s cross-page context holds up on real-world files.
- Hardware and memory requirements for running it through Transformers or SGLang.
Discussion spark: For long-document AI, would you prioritise a model that can process more pages at once, or one that is demonstrably more accurate on the pages it does handle?
Sources and evidence
- Unlimited OCR Works 2 (27 October 2026, 13:44 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.