ATH-MaaS released OvisOCR2, a 0.8B vision-language model post-trained from Qwen3.5-0.8B that serves as the first end-to-end model to top the OmniDocBench leaderboard with a score of 96.58 on v1.6.
- It generates complete markdown including HTML tables and LaTeX formulas from single page images, eliminating the layout detection failures common in pipeline OCR systems.
- In tests on 827 real scanned medical documents, OvisOCR2 achieved a mean F1 of 0.947 compared to GLM-OCR's 0.908, with significant improvements in worst-case scenarios (p10 F1 of 0.898 vs 0.810).
- On a single RTX 5090 using vLLM 0.22.1, the model reaches peak throughput of 590 pages per minute at concurrency 32, with optimal performance at 150 DPI.
- The author notes that while greedy decoding can cause repetition loops in ~0.5% of cases, applying a frequency penalty on retry resolves the issue without distorting legitimate content.
The model offers higher accuracy and simpler deployment than traditional pipeline OCR, particularly for documents with complex layouts or poor scan quality.