LaunchAI ModelsOctober 27, 2026

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

Read original source →huggingface.co

The 3.34B vision-language model handles 131,072-token contexts and parses hundreds of pages in one request via SR-SWA sparse attention, which keeps per-step decoding cost nearly flat. Baidu reports 77.1 overall on MarathonDocBench versus 33.7 for Kimi-K3; weights ship under MIT.

2 sources

More stories today

Open the live feed