Baidu’s Unlimited-OCR Reads A 40-Page PDF In One Pass — Here’s What The Viral Posts Get Wrong, And What Actually Matters

Baidu has open-sourced Unlimited-OCR, a 3-billion-parameter model capable of processing entire multi-page PDFs in a single forward pass. This development marks a significant technical achievement in OCR technology, with potential implications for document processing and AI deployment, especially given its support for self-hosting and open licensing.

The model was released on June 22, 2026, with a technical report following on June 23. It is available on Hugging Face under an MIT license, supporting frameworks like Transformers, vLLM, SGLang, and Docker. Built upon DeepSeek-OCR, Unlimited-OCR introduces a novel Reference Sliding Window Attention (R-SWA) mechanism that replaces traditional linear cache growth in decoder models, enabling processing of dozens of pages in a single pass without external schedulers or page splitting.

According to the technical report, Unlimited-OCR achieves a throughput of approximately 5,580 tokens per second on OmniDocBench, outperforming its predecessor DeepSeek-OCR by about 12.7%. It scores over 93 on OmniDocBench v1.5, positioning it at the top of end-to-end document parsing benchmarks. Its key advantage lies in its ability to parse long documents—up to 40 pages—with an error rate below 0.11, a feat not matched by page-by-page models, which struggle with cross-references and reading order.

However, claims circulating about 1.9 million downloads are inaccurate; the model page reports around 8,400 downloads in the last month, indicating high but not viral-scale adoption. Notably, Baidu’s own PaddleOCR-VL and Zhipu’s GLM-OCR models outperform Unlimited-OCR in peak accuracy, but they evaluate page-by-page, not in a single multi-page pass.

Implications of Constant Memory OCR for Long Documents

This development demonstrates a practical approach to overcoming the memory limitations of decoder-based OCR models, enabling more efficient processing of lengthy documents without splitting or stitching. It challenges the reliance on cloud-based OCR solutions by providing a self-hosted, open-source alternative capable of handling complex, multi-page content with high speed and accuracy. While it may not surpass all models in accuracy on single pages, its architecture offers a new paradigm for long-form document analysis, with potential impacts on legal, academic, and enterprise workflows.

Technical Roots and Prior OCR Advancements

Unlimited-OCR is based on DeepSeek-OCR, which itself integrated innovations like SAM-ViT and CLIP-ViT for visual encoding. The key architectural change introduced by Baidu is the Reference Sliding Window Attention (R-SWA), which mimics human-like ‘soft forgetting’ and maintains a fixed memory footprint regardless of output length. This approach addresses a longstanding challenge in decoder-based OCR models, where memory and latency grow linearly with the number of tokens generated.

Previous models such as PaddleOCR and Zhipu’s GLM-OCR have achieved higher peak accuracy but rely on page-by-page processing, limiting their effectiveness for long, interconnected documents. The breakthrough with Unlimited-OCR is its ability to process entire multi-page PDFs in a single pass, maintaining low latency and stable memory consumption, which is a significant step forward in AI-powered document analysis.

“Unlimited-OCR introduces a novel attention mechanism that enables processing of dozens of pages in a single forward pass, with constant memory usage.”

— Baidu Research Team

Unconfirmed Claims and Limitations of the Model

While the technical achievements are clear, some claims about the model’s performance—such as being the ‘most accurate’ or the ‘fastest’—are relative and depend on benchmarks. The report notes that Unlimited-OCR is not the highest-scoring model on all benchmarks, with models like PaddleOCR-VL and Zhipu’s GLM-OCR outperforming it in peak accuracy, but they evaluate pages individually rather than in a single pass. Additionally, the true impact on real-world workflows remains to be tested outside controlled benchmarks.

Further, the actual adoption rate is much lower than viral figures suggest, with only around 8,400 downloads in the last month, indicating niche but growing interest. It is also unclear how well the model performs on diverse, real-world documents outside the technical benchmarks.

Next Steps for Adoption and Benchmarking

Further independent testing and real-world deployment will clarify the model’s practical advantages. Baidu is expected to continue refining the architecture and possibly release more optimized versions or integrations. The open-source community may adapt the model for various applications, from legal document analysis to academic research. Monitoring its adoption and benchmarking against emerging models will be key to understanding its long-term impact.

Key Questions

How does Unlimited-OCR process entire multi-page PDFs in one pass?

It uses a novel attention mechanism called Reference Sliding Window Attention (R-SWA), which maintains a fixed-size memory cache, allowing the model to read and analyze multiple pages without increasing memory or latency linearly.

Is Unlimited-OCR the most accurate OCR model available?

No, models like PaddleOCR-VL and Zhipu’s GLM-OCR achieve higher peak accuracy on page-by-page benchmarks. Unlimited-OCR’s strength is in processing long documents in a single pass, trading some accuracy for efficiency and memory stability.

Can I run Unlimited-OCR on my own hardware?

Yes, the model is open-source, MIT-licensed, and supports frameworks like Transformers, vLLM, and Docker, making it accessible for self-hosted deployment.

What are the limitations of Unlimited-OCR?

Its performance on non-benchmark, real-world documents remains to be fully tested. Also, it does not outperform all models in peak accuracy, focusing instead on long-document processing efficiency.

How does this impact the OCR industry?

This development introduces a new architectural approach that could shift how long documents are processed, reducing reliance on external OCR services and enabling more efficient, self-hosted solutions for complex workflows.

Source: ThorstenMeyerAI.com