Unlimited OCR 2
Baidu open-sourced Unlimited OCR 2, a document parsing model that reads documents of any length in one pass, scoring 84.6 on MPDocBench and 77.1 on MarathonDocBench.
Version 2 replaces the reference sliding window attention of the June original with Sparse Reference Sliding Window Attention (SR-SWA). A page gate narrows each query's visual reference to relevant pages, and top-scoring buckets of visual tokens are pulled in, so the working set stays fixed (for example 512 tokens) and decoding latency and memory stay nearly constant as documents grow. Baidu says it parsed all seven Harry Potter books, over 3,600 pages, in one run.
- Date
- Wednesday 28 October 2026
- Lab
- Baidu
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| MPDocBench score | 84.6 multi-page parsing benchmark | company |
| MarathonDocBench score | 77.1 hundred-page document benchmark released by Baidu alongside the model | company |
| Decoding speed per step at 100 pages vs full attention | 7x faster SR-SWA; 5.6x faster than previous R-SWA | company |
Weights and code are under the MIT license on GitHub and Hugging Face, and the model is also on Baidu AI Cloud. Benchmark figures are Baidu's own and no independent evaluations had been published.
Sources
This record was checked against its sources on 28 October 2026. How we check