Baidu Unlimited OCR: Claims and Verifications for a Local Long-Document OCR Tool

On October 11, 2026, the X account Brooks Whale X (@BrooksWhaleX) published an introduction to Unlimited OCR, linking to a Hugging Face repository under the Baidu name. The author presents it as a locally runnable long-document OCR tool, with the repository link supplied in a follow-up post.

This article summarizes only what the author's posts claim. Figures and performance statements have not been independently verified.

Repository, Model Scale, and License Claims

The repository identifier supplied in the author's follow-up post is baidu/Unlimited-OCR, at https://huggingface.co/baidu/Unlimited-OCR. The bundle contains no direct inspection of the repository page, so only that the identifier URL carries the baidu organization name is confirmed.

On model scale, the author's reply claims roughly a 3.3B MoE model with about 500M parameters active per token. The same reply claims the model is built on the DeepSeek OCR architecture. The author also claims an MIT license.

Single-Pass Long-Document Claim and Supported Runtimes

The author claims multi-page documents are processed in a single forward pass within a 32K context window. The stated mechanism, from the author's reply, is R-SWA attention keeping the KV cache constant plus heavy visual token compression in the encoder. The same reply adds that real-world results depend on document complexity, layout, and hardware.

For local execution and deployment, the author's root post claims support for Transformers, vLLM, SGLang, Docker, Ollama, and llama.cpp, alongside 100% local execution. Multilingual support is likewise claimed in the root post.

Numerical Claims and Pre-Adoption Checks

The root post claims 93% on a standard parsing benchmark, six points above baseline, and an error rate under 0.11 past 40 pages. The benchmark name and comparison baseline are not specified in the post. These are the author's claims; the bundle contains no independent verification material.

The cost comparison is also the author's claim. The author contrasts per-1,000-page costs of $1.50 to $15 for Textract, Google Vision, and Azure Doc Intelligence against free local execution. Because this is unverified, check each service's current pricing before citing it as a rate comparison.

Cautions for adoption decisions follow from the evidence. The root post advertises 100-page single-pass reading, but the only page-anchored numerical claim is the past-40-pages error-rate figure, so it cannot be read as proven 100-page performance. One thread reply notes the model was released a few months earlier, so framing it as a fresh open-source release is inaccurate on the bundle's evidence. The author's own reply conditions results on document complexity, layout, and hardware, so teams should test it on their own documents before production use.

Sources

• Baidu Unlimited OCR repository: baidu/Unlimited-OCR
• Brooks Whale X (@BrooksWhaleX) original thread: October 11, 2026 introductory post