Baidu’s Unlimited-OCR Reads Dozens of Pages in One Pass

OCR is hot in Korea, and multipage jobs still break the same way: split pages, OCR each, glue results, lose context and formatting.

Baidu released Unlimited-OCR and shook that workflow again.

1. ~500M active parameters yet 93.23% on OmniDocBench v1.5, above 235B Qwen3-VL (89.15%) and Gemini-2.5 Pro (88.03%). Architecture beat parameter bragging. (HF)

2. R-SWA (Reference Sliding Window Attention) keeps image tokens and prompts looking at all pages while generation only attends the last 128 tokens of output. KV cache becomes a fixed-size queue. 10k or 100k generated tokens, same memory footprint. (GitHub)

3. 40+ pages without the classic slowdown. Old models grow KV with pages and decoding collapses. Fixed cache size removes that bottleneck on contracts and long reports.

4. Workflow change: before = split → per-page OCR → external merge. After = normalize format → whole PDF in one call. Cross-page structure stays intact.

5. infer_multi + PyMuPDF to images, optional sglang OpenAI-compatible serve. MIT license, commercial use open.

OCR is moving from “read characters” to “understand multipage documents as one engine.” If you run document pipelines, study R-SWA.