Unlimited-OCR-2
modelYour notes
The second Unlimited-OCR, released October 27, 2026 under MIT, which moves Baidu's one-shot document parser from long documents to documents of effectively any length. The change is in attention: every decoder attention module becomes Sparse Reference Sliding Window Attention (SR-SWA), which caps the text-to-image references each decoding step can attend to at a fixed set (512 tokens in the card's example). Per-step latency and memory therefore stay roughly constant as the document grows, instead of rising with the accumulated visual context. At the hundred-page scale the paper (posted October 29) reports 7.0× faster decoding steps than full attention and 5.6× faster than the first model's R-SWA. It claims state-of-the-art end-to-end results on MPDocBench and on MarathonDocBench, a benchmark Baidu released alongside it: 100 documents of 100–300 pages each, about 18K pages and 205K annotated layout elements in mixed Chinese and English, with cross-page text and tables. The paper frames this as the start of an "OCR 3.0 era". The HuggingFace repo drew about 570 likes in its first week.