Baidu's "Unlimited OCR" processes dozens of document pages in one pass by treating memory like human forgetting

2026-07-05 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, medium

Summary

Baidu researchers have developed "Unlimited OCR," an optical character recognition model capable of processing dozens of document pages in a single inference pass while maintaining constant memory usage and speed, irrespective of text length. This is achieved through Reference Sliding Window Attention (R-SWA), a redesigned mechanism that allows generated tokens to attend to all reference tokens but only the last 128 output tokens, fixing the KV cache size. Built upon the open-source Deepseek OCR, Unlimited OCR integrates its DeepEncoder with a 3-billion-parameter Mixture-of-Experts decoder. The model scores 93% overall on OmniDocBench v1.5, a six percentage point improvement over Deepseek OCR, and 93.92% on v1.6, leading end-to-end system rankings. It sustains an error rate below 0.11 beyond 40 pages and boosts processing speed by 12.7% to 5,580 tokens per second in Base mode.

Key takeaway

For Machine Learning Engineers deploying OCR solutions for multi-page documents, Baidu's "Unlimited OCR" offers a significant advantage. You can now process dozens of pages in a single pass, achieving constant memory use and speed, unlike traditional methods. This eliminates the need for page-by-page processing loops, boosting throughput by 12.7% and improving accuracy on benchmarks. Consider integrating this model or its R-SWA mechanism to enhance your long-document parsing capabilities.

Key insights

Unlimited OCR uses fixed-window attention to process multi-page documents efficiently, mimicking human selective recall.

Principles

• Fixed-length KV cache prevents memory growth.
• Exempting visual tokens from state changes preserves image quality.
• Limited attention can improve focus on dense OCR tasks.

Method

The DeepEncoder compresses images to 256 tokens, then an MoE decoder with R-SWA processes them, running the KV cache as a fixed-length queue for the last 128 tokens.

In practice

• Integrate Unlimited OCR with vLLM or SGLang for inference.
• Explore R-SWA for other reference-based tasks like speech recognition.
• Utilize the model for efficient multi-page document parsing.

Topics

• Optical Character Recognition
• Reference Sliding Window Attention
• KV Cache Optimization
• Long Document Parsing
• Deepseek OCR
• Mixture-of-Experts

Code references

• baidu/Unlimited-OCR

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Unlimited OCR in 6 mins! — Baidu's Unlimited OCR leverages reference sliding window attention for high-throughput, memory-efficient document parsing.
• WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering — WaveFilter uses wavelet transform to identify critical tokens in long sequences, creating a sparse KV Cache to improve Diffusion LLM performance.
• How Attention Got So Efficient [GQA/MLA/DSA] — Efficient attention mechanisms like DSA significantly reduce LLM compute costs and memory while preserving performance.
• Multilingual OCR-Aware Fine-Tuning and Prompt-Guided Chain-of-Thought Reasoning for Multimodal Large Language Models — OCR-aware data curation and fine-tuning significantly improve MLLM robustness for multilingual text in degraded images.
• Fast Large Table Extraction: Sparrow + dots.ocr to JSON — Sparrow's new dots OCR integration significantly accelerates large table processing by offloading structured data extraction to custom logic.
• OCR vs. Vision LLMs: Choosing the Right Tool for Intelligent Document Processing — The optimal Intelligent Document Processing strategy combines traditional OCR with Vision LLMs based on document complexity, volume, and cost.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.