Unlimited OCR, Limited Reach

The most-hyped open OCR model of 2026 is not the best one. Baidu's Unlimited-OCR made a splash by reading an entire multi-page PDF in a single pass. This episode unpacks what that one-shot trick costs, and who should actually use it.

We explain the constant-memory attention design that keeps throughput flat across dozens of pages, then look at the catch. To fit a whole document into one pass, the model drops to a downscaled Base mode of about 256 visual tokens per page, and fine footnotes, dense tables and small type start to slip. We walk the scoreboard on OmniDocBench and olmOCR-Bench, where PaddleOCR-VL, olmOCR-2, DeepSeek-OCR 2 and MonkeyOCRv2 each win a slice. We examine the disputed 5,580 tokens-per-second claim, which is Baidu's own single-source figure, and the flat decoding latency that does hold up. Then we cover the deployment reality: dedicated Docker images, a mandatory custom logits processor, and no GGUF or llama.cpp support.

The verdict is that Unlimited-OCR is a specialist engine for long, unchunked, text-heavy documents, not a general default. If you run a document pipeline, this episode tells you when to reach for it, and when to reach for something else.