Baidu's Unlimited-OCR paper reports 93.23, but 32K caps real use

Unlimited-OCR is the newest name in a document-OCR field that already includes PaddleOCR, GLM-OCR, DeepSeek-OCR and MinerU. The model's origin is specific: it starts from DeepSeek-OCR and replaces the decoder's attention with Reference Sliding Window Attention (R-SWA), a design that limits how much of the recent generated output the KV cache must keep. Baidu published the code and weights, and the paper was first posted to arXiv on June 22, 2026.

What that design does and does not mean is easy to lose in the name. Keeping the KV cache from growing with output length is not the same as saying GPU memory stays flat as pages are added, because the input image, the model weights and preprocessing all still consume memory. Likewise, the paper's internal evaluation includes documents over 40 pages, but the standard maximum length is 32K, and the official multi-page inference example specifies a maximum of 32,768. The cloud service's 500-page API limit applies to the submitted document, not to what the local model can read in one pass, and the update date on the API documentation is not the model's release date.

The choice between running the model locally and calling Baidu Cloud's API is not a choice between the same thing twice: the API is asynchronous — you submit a task, receive an ID, then fetch the result — while the local infer_multi call should not be assumed to share the same input rules or limits. Baidu Cloud's current documentation offers a limited-time free tier and free-test allowances of 200 pages for individual real-name verification and 1,000 pages for corporate verification.

The accuracy comparison rewards careful reading as well. The paper's Table 1 restricts itself to end-to-end VLM models; treating 93.23 as a win over every OCR system would stretch that comparison past what the table covers, since configurations such as GLM-OCR combine layout analysis with region recognition. The evaluator's published v1.6_full table is a different set of numbers from a different source, and the paper's own v1.5 figures and self-reported v1.6 value of 93.92 are not mixed with v1.6_full entries such as 94.00. Version labels matter within a family too: MinerU2.5-Pro and MinerU-2.5 refer to different evaluated checkpoints (MinerU2.5-Pro-2605-1.2B and MinerU2.5-2509-1.2B). For a Japanese-language study use case, the v1.6_full table alone does not settle rankings for vertical writing, ruby, or handwritten answers, since the evaluation attributes center on English, Simplified Chinese, and mixes of the two — which is why the recommended next step is not another benchmark but a small, purpose-built test set of one's own documents.