OvisOCR2 (0.8B): first end-to-end model to top OmniDocBench - I threw 827 real scanned medical docs at it, here's everything I learned

What it is: ATH-MaaS/OvisOCR2 - a 0.8B document-parsing VLM post-trained from Qwen3.5-0.8B (SFT + RL + OPD), Apache 2.0, runs on vLLM 0.22.1. One prompt per page image -> complete markdown (HTML tables, LaTeX formulas, bbox stubs for figures). It scores 96.58 on OmniDocBench v1.6, and it's the first