Baidu's open Unlimited-OCR model parses long documents at flat memory cost, beating DeepSeek's baseline

Baidu’s Unlimited-OCR, a 3-billion-parameter open model for parsing documents, has drawn a surge of developer interest since Baidu released it under an MIT license on June 23, 2026. The appeal is an architecture built to read long documents without running out of memory.

Unlimited-OCR is a Mixture-of-Experts model that activates only 500 million of its 3 billion parameters at inference, keeping it cheap to run. Its central trick is what Baidu calls Reference Sliding Window Attention, which holds the key-value cache flat regardless of document length. That lets the model parse multi-page PDFs and images in one pass, without the memory cost that usually climbs as input grows.

Baidu reports strong accuracy. On the OmniDocBench v1.5 benchmark, Unlimited-OCR scores 93.23, which the company says beats a DeepSeek OCR baseline by 6.22 points. The results are Baidu’s own and have not been independently verified. They are also uneven: on the ParseBench suite the model posts a mean of 46.17 and a Text Content score of 86.81, but a Text Formatting score of just 0.97, a weak spot the model card discloses.

The model runs through Hugging Face Transformers, SGLang and vLLM, and ships in two image-resolution settings, a 640-pixel ‘gundam’ mode and a 1024-pixel ‘base’ mode, per Baidu’s code release. Document parsing has become a competitive corner of open AI because it feeds retrieval and agent pipelines that need to turn messy files into structured text.

Whether Unlimited-OCR’s flat-memory approach holds up on independent document sets, or its low formatting score limits real use, will determine how far the early enthusiasm carries.