Unlimited-OCR-2
Unlimited-OCR-2 is a document parsing model released on 27 October 2026. Sparse-reference sliding-window attention keeps a limited set of working references while processing long documents.
Model design and access
The official implementation has separate single-page and multi-page inference paths. Multi-page inference places pages in one query, while the single-page path can use dynamic image tiles. The card links a Baidu Cloud OCR service and provides local inference instructions. Its long-document design does not establish a numerical infinite context window.Unlimited-OCR-2 official model card
Practical use
Practical guidance. Evaluate page continuity as well as character accuracy. Long invoices and reports can lose table headers, footnotes or references between pages even when isolated text is correct. Check the output against the page images before using it as structured source data.Unlimited-OCR-2 official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable, API |
| License | MITUnlimited-OCR-2 official model card |
| Input | Single-page and multi-page document imagesUnlimited-OCR-2 official model card |
Related models
- Unlimited-OCR
Baidu · Unlimited-OCR
DeepSeek-OCR-based model for joint parsing of multiple pages in a single document-level pass.
Launch date Downloadable - Qianfan-OCR
Baidu · Qianfan-OCR
Four-billion-parameter document model for recognition, layout analysis, question answering and field extraction in 192 languages.
Launch date not confirmedDownloadable