Baidu released Unlimited OCR 2 on October 27, 2026, according to its dated model-card changelog. Alongside it, OpenBMB’s available MiniCPM-OCR 1.0 and MiniCPM-OCR Jev checkpoints illustrate two other parts of document automation: extracting content and checking candidate OCR results.
Unlimited OCR 2 focuses on long-document continuity
The Baidu release introduces Sparse Reference Sliding Window Attention, or SR-SWA, to reduce the attention state carried through long parsing sequences. Its documentation provides multi-page inference that preserves page order and uses cross-page context. The checkpoint is marked MIT licensed.
Baidu calls this unlimited-horizon parsing, but the name should not be interpreted as an absence of operational limits. Its serving documentation still includes page windows, output budgets and memory controls. It explicitly notes a tradeoff: smaller windows reduce peak memory but can weaken merging across page boundaries.
That makes document structure part of capacity planning. A table that continues onto the next page is different from a stack of unrelated receipts, even when both inputs contain the same number of pages.
MiniCPM-OCR 1.0 handles the extraction stage
OpenBMB’s 1.3B parser targets Chinese and English documents, including scans and photographs, and covers text, tables, formulas, charts and seals. The model card labels it Apache 2.0. The project repository provides deployment documentation and links to related accelerated and quantized versions.
Its role is to produce document content in a useful form. A small parameter count can make deployment more approachable, but it does not establish acceptable accuracy for every layout, handwriting style or business field. Vendor benchmark rankings are separate from a team’s own document acceptance criteria.
Jev assesses a candidate rather than replacing the parser
MiniCPM-OCR Jev is a 2B correctness decision model. It takes a text, formula or table region together with a candidate OCR result and returns a score between zero and one. Its Apache 2.0 card describes using that score to select among candidates or route uncertain regions for review.
The score is not an automatic guarantee that a region is correct. A deployment must establish how the scores relate to its own error tolerance. The October 27 Baidu changelog supplies the dated release anchor here; the OpenBMB models are related available checkpoints, not an assertion that all three launched simultaneously.
Build the review loop around consequential fields
Our assessment is that the practical opportunity lies in combining extraction with targeted review. For invoices, the highest-cost error may be a single digit in an account number or total. A page that mostly looks right can still contain an unacceptable payment instruction.
A proposed evaluation should preserve original pages and region coordinates, then compare extracted fields with a human-checked sample. Include long tables, repeated headers, skewed photographs and low-quality scans. Track corrections by field and document type, rather than only averaging text accuracy across the batch.
Introduce automatic acceptance only after measuring false acceptance at the chosen review threshold. Keep a sample of high-scoring regions in the human queue to detect drift, and make the original image accessible to the reviewer.
These releases address complementary bottlenecks. Long-context parsing helps maintain continuity; document extraction produces usable content; a correctness model helps prioritize inspection. None removes the need to prove that the final records match the source before consequential downstream writes.