Baidu releases “Unlimited OCR 2,” a new open-source OCR model designed to parse documents of very great length in a single pass. The system is built to keep decoding cost nearly constant as document size increases, using a sparse reference sliding window attention approach to limit how attention scales during processing.

The model is released with weights and code made available on major model-hosting platforms, under the MIT license. One outlet emphasizes the capability to handle extremely long inputs—describing use cases such as hundred-page documents parsed in one request—while another highlights the technical mechanism and the implementation availability. Both accounts align on the core release details (open sourcing, MIT licensing, and the one-pass design) and on the attention technique aimed at controlling computational cost as documents grow.