Baidu Releases Unlimited OCR 2 for Long Documents

Baidu has expanded its artificial intelligence portfolio with the release of Unlimited OCR 2, an MIT-licensed optical character recognition model boasting 3 billion parameters. This new system is engineered specifically to overcome traditional bottlenecks when dealing with exceptionally lengthy documents spanning hundreds of pages in a single workflow.

Processing massive files has historically caused computational overheads to scale dramatically with page counts. However, Baidu's latest architecture effectively bypasses this barrier, allowing systems to ingest and parse extensive dossiers simultaneously without requiring tedious manual segmentation or piecemeal processing.

Stock photo for illustration only, not from the actual event

The core innovation driving this performance is its SR-SWA attention mechanism. This specialized approach ensures that the per-step decoding cost remains nearly flat, regardless of how much additional length is introduced by the source document, thereby maintaining steady execution speeds and stable resource consumption.

Demonstrating its capability under rigorous testing, Unlimited OCR 2 achieved a score of 77.1 on MarathonDocBench, reinforcing its precision and competitive standing for complex, multi-page text extraction tasks.

Source: MarkTechPost