Baidu open-sourced a 3B model that reads a 40-plus-page PDF in a single pass — no page-by-page stitching, no lost context — and it runs locally on one 8GB GPU.
Unlimited-OCR replaces standard decoder attention with Reference Sliding Window Attention (R-SWA), which keeps the memory footprint flat no matter how long the document is, so it transcribes a whole multi-page file in one forward pass instead of page by page. It scores 93.23 on OmniDocBench v1.5 — about 6 points above the prior best — preserves layout, and outputs tables as HTML and equations as LaTeX. It is MIT-licensed and runs on vLLM, SGLang, Ollama, and Transformers.
For multi-page trade documents, a local, zero-API-cost model that holds context across the whole file addresses the main weakness of page-by-page OCR — lost context between pages.
Customs packets, bills of lading, rate sheets, and commercial invoices are long, multi-page, and table-heavy — the exact case where page-by-page OCR drops context. A model that reads the whole document at once, keeps layout, and runs on-prem with no data leaving the building is a direct fit for freight back-office automation.
https://www.marktechpost.com/2026/06/24/baidu-releases-unlimited-ocr-a-3b-model-that-keeps-the-kv-cache-flat-for-long-document-parsing/