Baidu researchers have developed an OCR model called 'Unlimited OCR' that processes multiple document pages in a single inference pass without increasing memory use or slowing down. This is achieved through a redesigned attention mechanism that mimics human forgetting, allowing the model to maintain constant speed and memory consumption regardless of the text length.

Current OCR models typically handle only about ten pages in a single pass, according to Baidu's technical report. The main bottleneck is the key-value (KV) cache, which limits the amount of information the model can retain during inference. Baidu's approach overcomes this limitation by rethinking how attention is applied across long sequences of text.

Unlimited OCR is built on top of the open-source Deepseek OCR model, which reduces visual tokens by 80% compared to earlier versions. This optimization helps improve efficiency and performance, making it possible for the model to process larger documents without significant increases in computational resources.

The ability to handle long documents in a single pass has implications for cost, vendor lock-in, and governance in AI systems. Organizations may reduce infrastructure costs by using fewer inference passes, but they may also become more dependent on specific models and platforms. Governance considerations, such as data privacy and compliance, become more complex as models process larger volumes of information in one go.

Baidu's Unlimited OCR is still under development, with ongoing research and improvements expected. The model's performance on tasks like document parsing and text recognition may continue to evolve, potentially influencing how OCR is used in industries such as finance, legal, and healthcare.