Baidu releases Unlimited OCR 2 with SR-SWA for unbounded-length document parsing
pmttyji · reddit · 2026-10-27
Baidu open-sourced Unlimited OCR 2, targeting unlimited-horizon document parsing via a new technique called Sparse Reference Sliding Window Attention (SR-SWA).
- SR-SWA compresses text-visual attention interactions over ultra-long sequences into a fixed-size set of working references (e.g., 512 tokens), keeping per-step decoding latency and memory nearly constant regardless of document length;
- This enables true unbounded document parsing within a finite context window, eliminating quadratic attention scaling and memory bottlenecks;
- Achieves SOTA on both dozens-of-pages benchmarks (MPDocBench) and extreme hundred-page benchmarks (MarathonDocBench).
The model is available on Hugging Face and GitHub.
Related event: Baidu Open-Sources Unlimited OCR 2 for Thousand-Page Document Parsing(21 posts)→
More from Models
- Cohere's Qwen sandbox incident sparks pushback: "poorly configured post-training artifact," says deliprao — PMinervini · 2026-10-27
- DeepSeek Flash impresses dev: let it build its own tools, cheap enough to experiment — victormustar · 2026-10-27
- Reddit asks: Why do almost all major open-weight models come from China? — PhilosopherSweaty826 · 2026-10-27
- Doctor's Prescription OCR Test: Gemma 4 31B Now Reads What Models Couldn't a Year Ago — MaziyarPanahi · 2026-10-27
- cyb3rops: open-weight AI models mean attackers can compromise 90% of orgs without new exploits — cyb3rops · 2026-10-27
- Study Finds GPT-6 Astra Only Partially Capable at CAD Workflows — ZhejiangUniversity · 2026-10-27