Baidu releases Unlimited OCR 2 for hundred-page document parsing

Baidu says Unlimited OCR 2 was released on Oct. 27, 2026, with a focus on parsing hundred-page documents.[1][2] The same day, Baidu released MarathonDocBench, which it describes as a benchmark for ultra-long document parsing.[1] The announcement links the model and benchmark around long-document workloads, giving teams evaluating document-processing models a specific task length and benchmark result to consider.[2] The release information identifies SR-SWA as the method through which the model achieves what Baidu calls unlimited-horizon parsing.[1]

The Hugging Face model card lists Unlimited OCR 2 as a 3-billion-parameter image-to-text model built with Transformers, with an MIT license.[1] These details can help technical teams identify the model type, framework and stated license as they decide whether to test it in their own document workflows. The card also makes the model available for download and use through Hugging Face.[1] Baidu says the model supports vLLM inference, another stated option for teams assessing how to run it.[1]

Baidu says Unlimited OCR 2 became available on ModelScope and Baidu Cloud on Oct. 27, 2026.[1] It also announced a demo on Hugging Face Spaces that day.[1] Baidu said its paper on the model had become available on arXiv on June 23, 2026.[1] The Hugging Face model page offers a download, while the listed deployment and demo options give teams several named routes to explore the model.[1] The release information also credits the ms-swift community with supporting training on Oct. 28, 2026.[1]

MarkTechPost reports that Unlimited OCR 2 scored 77.1 on MarathonDocBench for hundred-page document parsing.[2] The benchmark is described by Baidu as covering ultra-long document parsing, and the model is presented for hundred-page documents.[1][2] For professionals, the score is a starting point for deciding whether to test the model on representative material; the available facts here do not establish how it performs on a particular organization’s documents. The model card’s stated size, license and framework, together with Baidu’s vLLM support claim, offer additional points to check during an evaluation.[1]