Model release · Baidu
Unlimited OCR 2
- Released
- Review updated
Baidu open-sourced Unlimited OCR 2 for document parsing. Its SR-SWA mechanism condenses text-visual attention into a fixed set of working references, so it parses documents of almost any length with near-constant decoding latency and memory. Baidu reports SOTA on MPDocBench and MarathonDocBench.
Availability
Open weights under the MIT license, about 3.3B parameters; Hugging Face baidu/Unlimited-OCR-2, code at GitHub baidu/Unlimited-OCR.