Unlimited OCR — Long-Horizon Document Parsing

Innovation

Baidu’s Unlimited OCR uses Reference Sliding Window Attention (R-SWA) replacing all standard attention layers in the decoder. This maintains a constant KV cache throughout decoding, solving the memory growth problem in long-document OCR.

Key Technical Details

• Based on DeepSeek-OCR encoder with high compression rate
• Constant KV cache regardless of output length
• Transcribes dozens of pages in a single forward pass at 32K context
• Two inference modes: “gundam” (640px crop) for single images, “base” (1024px) for full-page/PDF
• General-purpose mechanism — applicable to ASR, translation, etc.

Deployment

• HuggingFace Transformers (torch bfloat16)
• SGLang server with OpenAI-compatible API
• PyMuPDF for PDF-to-image conversion

Hermes Relevance

• Potential for document parsing pipeline improvement
• R-SWA constant KV cache concept may apply to agent context window management
• Strong open-source contribution from Baidu