Baidu announces Unlimited OCR 2

A dated release notice for document parsing

Baidu’s release log and contemporary PC Watch coverage date the announcement to October 27, when the project also listed a demo and MarathonDocBench. The repository contains an October 26 weight upload, so the announcement date should not be read as a verified first-public-download date. The pinned revision establishes the earlier files, without establishing when repository visibility changed. [1][2][3][4]

The model card identifies MIT licensing and describes Sparse Reference Sliding Window Attention. Baidu says it bounds the historical visual references used during decoding; the claim concerns per-step attention costs, rather than proving that the entire document-processing pipeline uses constant memory. [1][2]