Baidu Releases Unlimited OCR: Breakthrough in Long-Sequence OCR Performance

Summary:Baidu has unveiled the Unlimited OCR model, which introduces the Reference Sliding Window Attention (R-SWA) mechanism to significantly enhance the performance of long-sequence OCR tasks. Traditional OCR models relying on LLM decoders suffer from increasing memory consumption and slower generation due to KV cache accumulation with longer outputs. Unlimited OCR maintains a constant KV cache, reducing computational costs and enabling the transcription of dozens of document pages in a single forward

---

Technical Breakthrough and Core Mechanism

Baidu's Unlimited OCR model is designed to address the performance limitations of traditional OCR technologies when handling long-sequence text. Existing OCR models typically rely on Large Language Models (LLMs) as decoders, allowing them to leverage the prior distribution of language for improved recognition accuracy. However, as the output sequence grows, the accumulated KV cache (Key-Value Cache) leads to exponential increases in memory consumption and slower generation speeds, particularly noticeable in tasks involving dozens of pages of text.

Unlimited OCR addresses these issues through the following technical innovations:

1. Reference Sliding Window Attention (R-SWA): This mechanism replaces traditional attention layers with a sliding window approach that dynamically computes attention weights, maintaining a constant KV cache size. This reduces computational costs and ensures efficient processing of long sequences.

2. High-Compression Encoder: Unlimited OCR's encoder inherits the high compression rate of DeepSeek OCR, further optimizing memory usage efficiency.

3. Versatile Design: The R-SWA mechanism is not only applicable to OCR tasks but also to tasks like Automatic Speech Recognition (ASR) and machine translation, demonstrating its versatility.

Engineering Trade-offs and Performance

• Memory and Speed Balance: By maintaining a constant KV cache, Unlimited OCR avoids the exponential growth of memory consumption during long-sequence processing while maintaining high generation speeds.

• Computational Cost: The R-SWA mechanism significantly reduces the computational cost of attention calculations, making the model more practical for long-sequence tasks.

• Multi-Task Applicability: The model excels not only in OCR tasks but also in ASR and translation, proving its robust design for various applications.

Developer Implementation and Deployment Recommendations

1. Model Deployment: Developers can leverage Baidu's open-source code and model weights to quickly deploy the Unlimited OCR model.

2. Task Adaptation: For tasks requiring long-sequence processing, it is recommended to prioritize the use of the R-SWA mechanism to enhance efficiency and reduce resource consumption.

3. Multi-Task Extension: Developers can experiment with applying the R-SWA mechanism to other tasks that involve long-sequence processing, such as video captioning and audio analysis.

Conclusion

The release of Unlimited OCR marks a significant breakthrough in the field of OCR for long-sequence processing. Its innovative R-SWA mechanism not only improves model performance but also demonstrates versatility across multiple tasks, providing developers with a more efficient and reliable solution.

---

Source: Hugging Face Trending Papers (2026-06-22)

— END —

Tags: #Baidu Follow #OCR Follow #R-SWA Follow #Long-Sequence Processing Follow #Open-Source Model Follow