Baidu has published Unlimited OCR 2, an open-weight document-understanding checkpoint that targets one of the harder operational problems in OCR: keeping information from many pages usable without attention memory expanding with every new page. Its official Hugging Face repository dates the release to October 27, lists an MIT license and exposes Safetensors model artifacts. The announcement is a meaningful model release rather than an API-only preview, although teams should still pin a specific repository revision and inspect the full license and implementation before deployment.
The architectural change is Sparse Reference Sliding Window Attention, or SR-SWA. Baidu describes compressing the text-visual interactions of long documents into a bounded set of working references, giving 512 tokens as an illustrative setting. A moving window and recurrently useful references aim to avoid the quadratic scaling of full document attention. The word 'unlimited' is a design ambition about horizon handling inside finite working memory, not evidence that unlimited images, tokens, latency or GPU usage can be processed in practice.
The model card documents a multi-page inference method that accepts ordered page images and returns page-marked text, so it can use context across page boundaries. Baidu also reports performance on MPDocBench and its MarathonDocBench long-document evaluation. Those scores and the near-constant per-step resource claims come from the publisher; Open Model Weights has not independently reproduced them, nor do they guarantee reliable parsing of arbitrary scans, complex tables, low-quality photos or handwritten material.
Deployment details temper the headline. Baidu's reference Transformers stack names Python 3.10, specific library versions, an NVIDIA A800 with 80 GB memory and trusted custom repository code; its SGLang instructions use specialized serving flags and an adapted runtime path. Operators should review the remote-code security boundary, test long mixed-layout documents with ground truth, measure end-to-end cost and confirm the desired output format. This article reports the accessible release and its engineering approach without turning publisher benchmarks into an unsupported universal quality ranking.
What this changes for the evidence layer
Downloadable first-party model candidate, not automatically a field-verified OMW record. Pin model SHA, inspect Safetensors and MIT terms, execute custom code only after security review, and measure real document fidelity before intake.
Baidu official Unlimited-OCR-2 Hugging Face model card
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗
