Details
- Baidu released Unlimited OCR 2 as an open-source model on October 27, 2026.
- The system is designed to parse documents of virtually any length within a finite context window.
- Its Sparse Reference Sliding Window Attention mechanism condenses text-visual interactions into a fixed-size set of working references.
- This approach is intended to keep per-step decoding latency and memory use nearly constant as documents grow longer.
- Baidu reports state-of-the-art results on MPDocBench and MarathonDocBench, which evaluate documents ranging from dozens to hundreds of pages.
- The release is available through Baidu’s GitHub repository and Hugging Face model page.
- The design addresses the rising computational cost and memory bottlenecks associated with full attention on long visual documents.
Impact
Unlimited OCR 2 could lower the infrastructure cost of processing lengthy scanned documents and make long-horizon OCR more practical for developers. By releasing the model and implementation openly, Baidu also gives researchers a way to test an alternative to attention methods whose computational demands grow sharply with document length. Its competitive significance will depend on independently verified accuracy, licensing terms, and performance against other long-context document models.