On this page

Baidu Inc. (@Baidu_Inc) said on October 27, 2026 that it has open-sourced Unlimited OCR 2, a document-parsing model aimed at multi-page and ultra-long image-text inputs. Weights are on Hugging Face as baidu/Unlimited-OCR-2 (MIT; safetensors total about 3.34B, often rounded to ~3.3B), with a model-card Release dated 2026/10/27. Code and docs live in the GitHub repository baidu/Unlimited-OCR, also under MIT. For other recent open multimodal releases on DigIntoAI, see MiniCPM-V 4.7: Open 35B Multimodal Model That Runs on Phones.

What SR-SWA changes

The headline technique is Sparse Reference Sliding Window Attention (SR-SWA). Per Baidu’s model card and repository, SR-SWA compresses text-visual attention interactions from ultra-long sequences into a fixed-size set of working references (the docs cite an example of 512 tokens). That bounds computational complexity so per-step decoding latency and memory footprint stay nearly constant as document length grows, instead of scaling with the full sequence. Baidu frames this as moving from “long” parsing toward unlimited-horizon parsing inside a finite context window—without claiming that every real-world PDF will fit every GPU budget.

Vendor-reported long-document scores

Baidu reports state-of-the-art results on MPDocBench (dozens-of-pages style) and MarathonDocBench (extreme hundred-page style). Those figures are vendor-reported from Baidu’s announcement and README—not independent third-party re-runs.

Inference paths: Transformers and SGLang

The model card says Unlimited OCR 2 supports Transformers (native, for quick tests) and SGLang (recommended for serving). Multi-page flow uses infer_multi: pass page images in order so all pages of a document are parsed in one request, with cross-page context. Single-page infer is also documented. Serving notes cover page-window knobs for very long documents and an OpenAI-compatible chat endpoint under SGLang. Teams already comparing open vision-language stacks may also want Nemotron 3.5 Super VL and Bolt: Specs, License, Runs on 24GB.

Why it matters

Open, MIT-licensed long-document OCR remains a busy lane for finance PDFs, contracts, and multi-page reports. Unlimited OCR 2’s pitch is less “another OCR checkpoint” and more attention scaling that stays roughly flat as page count rises, plus a documented multi-page one-shot path. Availability is weights plus inference code on Hugging Face and GitHub as of the October 27, 2026 release note—not a claim that every enterprise stack should rip out prior OCR today. For another open visual/geometry release in recent coverage, see Hi3D 1.0 Is Open Source: Image-to-3D, Formerly Sparc3D.