Unlimited OCR 2: parse a whole PDF in one request. even on a 16 GB GPU

Unlimited OCR 2 parses a whole PDF in one request, and keeps track of text and tables that cross a page break. Turn its page tiles off and a 59-page tax guide comes back in 3 min 44 s on one RTX 4090, 4.2% off the PDF's own text, with the same quality on a 16 GB card. We deployed Baidu's MIT-licensed document parser (3B MoE, about 0.5B active parameters) on InferenceKey Cloud as a custom Python backend running the authors' own SGLang build, and sent it IRS Publication 15 (59 pages, public domain) in one streamed request. What we measured (one RTX 4090, 24 GB): - One view per page (crop_mode=false): 59 pages at 15.8 pages/min, first text after 3.3 s, 4.2% edit distance to the PDF's text over the whole guide - Authors' default page tiles: 8.7 pages/min, and the long table on pages 46 to 54 broke: eight rows squeezed into one cell, one sentence repeated 1,508 times, empty pages after it (38.1% overall, 2.4% away from the table) - Pages 46 to 54 alone: 28 s end to end with one view per page, each continuation page marked as part of the same table - One window for the whole document instead of 10-page windows: hit the 131,072-token context and stopped on page 27, 50.3 s to the first text - Memory: peak 19.4 GB on the 4090; a real 16 GB RTX 2000 Ada parsed the whole guide with the same quality (4.23% vs 4.18%) at 13.8 GB peak, 5.6 pages/min, about 2.8x slower Benchmark figures and the whole-novel replay in the video are the authors', not reproduced by us. Our numbers come from one document, not a benchmark. Full write-up: https://inferencekey.com/research/unlimited-ocr-2-whole-book-parsing SDK example: https://github.com/inferencekey/inferencekey-sdk/tree/main/examples/unlimited-ocr-2-pdf-parsing-custom-backend-private-nvidia Docs: https://docs.inferencekey.com 00:00 A long tax table, parsed in 28 s 00:18 What is Unlimited OCR 2? 01:09 Does a table survive the page break? 02:40 How small a card will do?