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 OCR model that can parse entire multi-page documents in a single forward pass, supported within a standard 32K context window. The model, released on June 22, 2026, aims to significantly improve long-document OCR efficiency, challenging the traditional page-by-page approach, and is available under an MIT license.

The model is based on Baidu’s DeepSeek-OCR architecture, incorporating a novel Reference Sliding Window Attention (R-SWA) mechanism that replaces linear cache growth with a constant-size cache. This innovation allows Unlimited-OCR to process dozens of pages simultaneously without increasing memory or latency, a breakthrough for long document OCR. The model’s technical report, published on June 23, 2026, details its architecture and performance metrics.

In benchmarks such as OmniDocBench, Unlimited-OCR scores 93.23 overall, outperforming DeepSeek-OCR’s 87.01, with notable improvements in text edit distance and table accuracy. It achieves a throughput of approximately 5,580 tokens per second, about 12.7% faster than its predecessor, especially on long outputs. For long documents, the model maintains low error rates across 20- and 40-page tests, with an edit distance below 0.11, though these results come from internal tests rather than independent leaderboards.

Contrary to viral claims, the model’s download figures on Hugging Face are around 8,400 in the last month, not 1.9 million, clarifying that its popularity is significant but not as exaggerated as some reports suggest. It is compatible with various frameworks, including Transformers and Docker, and supports community quantizations for deployment across different platforms.

Architectural Innovation in Long-Document OCR

Unlimited-OCR’s introduction of constant memory attention marks a notable advancement in OCR technology, enabling the processing of entire multi-page documents in a single pass. This reduces the need for splitting PDFs into pages, preserves reading order, and improves accuracy in tables and cross-references, which are common issues with traditional page-by-page methods. While it does not currently surpass the highest single-page accuracy models, its ability to handle long documents efficiently makes it a valuable tool for applications requiring comprehensive document understanding.

Baidu’s OCR Model Lineage and Recent Advances

Baidu’s OCR models have historically been based on architectures like PaddleOCR and DeepSeek-OCR. The recent release builds on this foundation, integrating a new attention mechanism to address memory growth issues inherent in decoder-based OCR models. The development aligns with broader industry efforts to improve long-form document processing, with competitors like PaddleOCR-VL and Zhipu’s GLM-OCR also making strides in accuracy. The release follows a series of incremental improvements, culminating in the ability to process entire multi-page PDFs in a single pass.

“Unlimited-OCR demonstrates that architectural innovation can enable true single-pass long-document OCR, with constant memory and faster throughput.”

— Baidu Research Team

Unconfirmed Claims and Limitations of the Model

It is not yet clear how Unlimited-OCR performs on independent, external benchmarks beyond internal tests. The reported low error rates for very long documents are based on proprietary datasets, not publicly available leaderboards. Additionally, claims of ‘state-of-the-art’ accuracy are nuanced; models like PaddleOCR-VL and Zhipu’s GLM-OCR outperform Unlimited-OCR on single-page benchmarks, though they lack the multi-page processing capability. The true real-world impact of the model remains to be validated across diverse document types and deployment scenarios.

Next Steps for Deployment and Industry Adoption

Baidu plans to further evaluate Unlimited-OCR on external datasets and benchmark it against industry standards. The open-source release invites developers to experiment with the model’s long-document processing capabilities and integrate it into applications requiring comprehensive OCR. Future updates may include fine-tuning for specific domains, improving accuracy on complex layouts, and expanding community-driven benchmarks to validate its performance in real-world settings. Industry adoption will depend on how well the model scales across different hardware and document types.

Key Questions

How does Unlimited-OCR differ from traditional OCR models?

Unlimited-OCR introduces a constant memory attention mechanism that allows it to process entire multi-page documents in a single pass, unlike traditional models that process pages individually and require stitching results.

What are the main performance benefits of Unlimited-OCR?

It offers faster throughput, especially on long documents, with a stable latency and fixed GPU memory usage, enabling efficient processing of large PDFs without splitting or external scheduling.

Can I run Unlimited-OCR on my own hardware?

Yes, the model is open-sourced under an MIT license and supports deployment via frameworks like Transformers, vLLM, and Docker, making it accessible for self-hosted use.

Does Unlimited-OCR outperform existing models in accuracy?

On certain benchmarks, it scores slightly lower than models like PaddleOCR-VL or Zhipu’s GLM-OCR in single-page accuracy, but it excels in processing entire multi-page documents efficiently.

Source: ThorstenMeyerAI.com