Baidu releases Unlimited-OCR-2 for unbounded document parsing
Original title: baidu/Unlimited-OCR-2 — new model trending #30 on Hugging Face
SR-SWA keeps a fixed working set, such as 512 tokens, and the release adds the MarathonDocBench benchmark.
Baidu published Unlimited-OCR-2, an OCR model using Sparse Reference Sliding Window Attention to parse documents of unbounded length inside a finite context window. The project says per-step latency and memory stay nearly constant as documents grow and claims state-of-the-art results on MPDocBench and MarathonDocBench. Weights are on Hugging Face and ModelScope, with Baidu Cloud availability, Transformers and SGLang inference, and a Hugging Face demo.
Unconfirmed: State-of-the-art results are the project’s own claims and are not independently verified here.
- MarathonDocBench, a hundred-page parsing benchmark, was released the same day.
- Multi-page inference parses every page in one request for cross-page context.
- Single-page mode uses dynamic tiling, called gundam mode, by default.
- The card lists Transformers 4.52.1 and a custom SGLang wheel for serving.
- Vendors
- Baidu
- Organizations
- ModelScope
- AI models
- Unlimited-OCR-2
From the original article
6 min read at huggingface.coimage-text-to-text · 513 downloads · 167 likes · license mit
By baidu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.