Baidu open-sources Unlimited OCR2, an OCR model that parses entire books in one request
abhishek__AI · x · 2026-11-23
Baidu's ERNIE team released Unlimited OCR2, an OCR model designed to read entire books — hundreds of pages — in a single request, breaking the usual document-length bottleneck.
- Traditional pipelines chunk PDFs and stitch them back; Unlimited OCR2 instead uses Sparse Reference Sliding Window Attention (SR-SWA) to keep a fixed-size working reference set (e.g. 512 tokens), keeping decoding latency and memory nearly constant
- Supports unlimited document parsing and multi-page one-shot parsing, targeting long PDFs, books and complex documents
- Ships with an OpenAI-compatible API, supports HuggingFace Transformers and SGLang; self-hosting requires no proprietary OCR API
- Reports SOTA on multiple benchmarks; the team demoed it parsing an entire Harry Potter book
Related event: Baidu Open-Sources Unlimited OCR2 for Book-Length Document Parsing(2 posts)→
More from Models
- Princeton's SSoN Benchmark: Best Vision Model Scores 39.7% vs Human 83.3% on Spatial Consistency — jiqizhixin · 2026-11-23
- Dev: 99% of people shouldn't self-host local LLMs, and he only does it for the hardware — gnukeith · 2026-11-23
- Theo says open models trail by 7.5 months; GLM-5.5 and Kimi-K3.5 may flip the story — scaling01 · 2026-11-23
- Opus 5.5 spots a missing factor of 2 in a 2015 paper's bound during math checking — xennygrimmato_ · 2026-11-23
- Musk Claims 25x Model Improvement, Not 25%; Ask: Will Distilled Models Benefit Too? — dreamwieber · 2026-11-23
- GLM-5.3-Flash 320B served from two Mac Studios: 137 tok/s single-user, 32 users, zero errors — AIFlow_ML · 2026-11-23