Unlimited OCR 2
Baidu's MIT-licensed 3B OCR model that parses multi-page and book-length documents with near-constant memory via SR-SWA attention.
Overview
Baidu released Unlimited OCR 2 on Hugging Face, with its release notes dating the release to October 27, 2026; the underlying paper (arXiv 2606.23050) appeared in June 2026. The 3B-parameter model uses Sparse Reference Sliding Window Attention (SR-SWA) so memory and per-step decoding cost stay nearly constant as documents grow, letting it parse multiple pages in one request, including book-length input. It accepts JPG and PNG page images and PDFs (converted at 300 dpi in the example code) and is MIT-licensed. Baidu claims state-of-the-art results on MPDocBench and the hundred-page MarathonDocBench but the model card gives no numeric scores.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 3B parameters, MIT license
- Sparse Reference Sliding Window Attention for near-constant memory on long documents
- Multi-page parsing in one request, demonstrated on a full-length book
- Accepts PDFs and JPG/PNG page images
Use cases
- Digitizing long reports, contracts and books in a single pass
- Self-hosted document parsing under a permissive license
Pricing
Free open weights (MIT license) on Hugging Face.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 27 Oct 2026
- Record updated
- 29 Oct 2026
- Record last checked
- 29 Oct 2026