What changed
Baidu released Unlimited-OCR with public weights and code for one-shot, long-horizon document parsing. The project later added a public demo, ModelScope artifacts and vLLM support while retaining the original 22 June release date.
Why it matters
Teams working with long reports can test an inspectable OCR stack whose memory design targets many-page inputs instead of sending every document to a closed hosted service.
What to do
Pin the repository and model revision, then compare page order, tables, handwriting, multilingual text, memory use and citation alignment against the current OCR pipeline on a private holdout set.
Caveat
Performance claims come from the authors. Long-horizon parsing can silently reorder or omit content, and the repository's later integrations should not be treated as evidence that the original checkpoint is accurate on a target document set.