All videos
0:00 / 0:00
research

Baidu Just Fixed DeepSeek-OCR for Reading Long Documents

The AI Automators30 June 2026Watch on YouTube

What you'll learn

  • Unlimited-OCR processes long documents in a single pass without slowdowns, handling dozens of pages at once
  • Optical compression (from DeepSeek-OCR) significantly reduces document image sizes on the input side
  • R-SWA (Reference Sliding Window Attention) maintains constant memory and speed regardless of how many pages are generated

Frequently asked questions

What makes Unlimited-OCR different from other OCR solutions?
Unlimited-OCR can process long documents in a single pass without slowdowns, which distinguishes it from traditional OCR models that require splitting documents. It builds on DeepSeek-OCR and uses optical compression and R-SWA attention.
How does R-SWA work and what is its advantage?
R-SWA (Reference Sliding Window Attention) is a flat KV cache mechanism that keeps memory and speed constant regardless of how many pages the model generates. This enables Unlimited-OCR to process long documents efficiently.
Is Unlimited-OCR the right tool for all document parsing tasks?
While Unlimited-OCR works well on long documents, there is a difference between something working and it being the right tool for the job, which depends on specific use cases and requirements.

Topics

ocrvision-language-modelsdocument-processingdeepseekattention-mechanisms

Read next

Description from the channel

👉 Access our AI Architects course & join hundreds of serious AI builders in our community: https://www.theaiautomators.com/?utm_source=youtube&utm_medium=video&utm_campaign=tutorial&utm_content=unlimited-ocr 🔗 Unlimited-OCR GitHub: https://github.com/baidu/Unlimited-OCR Hugging Face: https://huggingface.co/baidu/Unlimited-OCR Paper (arXiv): https://arxiv.org/abs/2606.23050 🔗 Other Resources DeepSeek-OCR (arXiv): https://arxiv.org/abs/2510.18234 Docling: https://github.com/docling-project/docling Datalab Marker: https://github.com/datalab-to/marker Last week Baidu open-sourced a new VLM called Unlimited-OCR, and the name is basically the whole pitch: take a long document and parse the entire thing in a single pass, dozens of pages at a time, without the model ever slowing down. It's already past 12,000 stars on GitHub, and it's a fine-tune of DeepSeek-OCR, the model that made a splash last year with optical compression. So I downloaded the open weights, got Claude Code to throw together a small app, and put it through its paces on a real 50-page PDF. It does work. But there's a big difference between something working and it actually being the right tool for the job. So in this video I get under the hood: the three kinds of OCR and when you'd reach for each, DeepSeek's optical compression for shrinking the input, and the one genuinely novel idea here on the output side, R-SWA (Reference Sliding Window Attention), the flat KV cache that keeps memory and speed constant no matter how many pages you generate. ⏱️ Timestamps: 0:00 Demo 03:01 Three kinds of OCR 06:57 Input Optical compression 08:50 Output flat-cache fix 11:56 Catches & verdict #AI #AIAgents #OCR #UnlimitedOCR #DeepSeekOCR #VLM #DocumentParsing #Baidu #Docling #KVCache #RAG #AgenticRAG #AIArchitects #AIBuilder