StudioBaidu releases Unlimited-OCR-2, enabling unlimited-horizon document parsing with SR-SWA
Original title: baidu/Unlimited-OCR-2 · Hugging Face
Baidu's team released Unlimited-OCR-2, which uses Sparse Reference Sliding Window Attention (SR-SWA) to compress text-visual attention over ultra-long sequences into a fixed set of working references (e.g., 512 tokens), keeping per-step decoding latency and memory nearly constant so documents of unbounded length can be parsed within a finite context window. The model reports SOTA on multi-page benchmarks like MPDocBench and extreme hundred-page benchmarks like MarathonDocBench; the model card provides Transformers and SGLang inference with memory-tuning knobs, plus vLLM inference and ms-swift training support. The paper is at arXiv 2606.23050.
The model card explains the SR-SWA design, long-document parsing results, and full Transformers/SGLang deployment configs, useful reference for whole-book document parsing workloads.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.