Baidu open-sources Unlimited OCR 2 for hundred-page document parsing

Baidu released Unlimited OCR 2, an MIT-licensed 3.34B-parameter OCR model that parses documents of hundreds of pages in a single request. Its SR-SWA sparse attention keeps decoding cost nearly flat, and it scores 77.1 on MarathonDocBench versus 33.7 for the next model.

• 3.34B parameters, MIT license, code on GitHub
• Max position length 131,072 tokens, runs on one CUDA GPU in BF16
• Scores 77.1 on MarathonDocBench; Kimi-K3 gets 33.7
• Weak spot is heading hierarchy at 48.2 (TEDS)