Unlimited-OCR 2 on a 16 GB GPU: matches Chandra 2, but "unlimited" fails

A blogger tested Baidu's Unlimited-OCR 2 (3.3B parameters, MIT license, released October 27) on a 16 GB RTX 4070 Ti SUPER. It ties Chandra 2 for top accuracy (0.12% vs 0.31% median character error rate), but ran out of memory at 80 pages after 74 minutes; with modified SGLang, per-token time stays flat.

• Unlimited-OCR 2: 3.3B parameters, MIT license, released by Baidu on October 27
• 0.12% character error rate vs 0.31% for Chandra 2 and 0.51% for Tesseract
• On 16 GB, memory rose from 7.5 to 11.1 GB at 10–40 pages; 80 pages hit OOM
• With SGLang, 4.17–4.19 ms per token, about nine times faster than transformers