Is Unlimited OCR 2 state of the art?Pending benchmarks.Announced Oct 27, 2026; no independent results yet.

The details so far

Lab

Baidu

Announced

Oct 27, 2026

API

Live · Baidu Cloud document parsing (Unlimited-OCR), an async task API based on Unlimited-OCR 2, free for a limited time

Open weights

Published · Hugging Face (baidu/Unlimited-OCR-2, Transformers with trust_remote_code) and ModelScope (PaddlePaddle/Unlimited-OCR-2); BF16

Parameters

3.34B (3,343,978,497 safetensors parameters)

Type

Vision-language OCR / document-parsing model (image-text-to-text)

Attention

Sparse Reference Sliding Window Attention (SR-SWA): text-visual attention compressed to a fixed-size set of working references (e.g. 512 tokens)

Multi-page parsing

infer_multi parses all pages of a document in one request with cross-page context, one global view per page; single pages use dynamic tiling (gundam mode)

License

MIT (model card and GitHub repo)

Serving

Transformers for quick tests; a customized SGLang wheel with an OpenAI-compatible API for serving

Benchmark released alongside

MarathonDocBench, a hundred-page document-parsing benchmark (baidu/MarathonDocBench on Hugging Face)

Predecessor

Unlimited-OCR (June 22, 2026), described in the arXiv paper 2606.23050; a separate model

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: leading results on long-document parsing benchmarks, MPDocBench (dozens of pages) and MarathonDocBench (hundreds of pages).

Lab claim: document parsing at virtually any length within a finite context window.

Sources