Baidu's Unlimited OCR reads dozens of pages in one pass by capping its KV cache
Baidu's Unlimited OCR parses dozens of document pages in a single pass by letting each generated token see every image token but only the last 128 tokens of its own output — Reference Sliding Window Attention — so the KV cache stays flat instead of growing with length. It scores 93% on OmniDocBench v1.5, six points above the DeepSeek OCR model it builds on, runs at 5,580 tokens per second, and holds error rates below 0.11 past 40 pages. That removes the page-by-page cache-reset loop current OCR systems depend on, and points at a wider use — compressing text into image tokens to stretch LLM memory across long documents.
Source: the-decoder.com ↗
No current OCR model handles more than about ten pages in a single pass, the Baidu researchers write in their technical report.
Why this matters
- → Processes 40+ document pages without resetting memory cache, enabling true end-to-end OCR.
- → Fixed KV cache removes computational bottleneck that forces current systems into page-by-page loops.
- → Opens path to compress images into tokens, extending LLM context windows for long documents.