A 0.8B OCR model beat our production pipeline on 827 scanned medical letters
OvisOCR2 — the first end-to-end model to top OmniDocBench — beat my production OCR pipeline on 827 real scanned medical letters and matched a frontier model at 0.8B parameters. The benchmark, the vLLM tuning that got one RTX 5090 to 590 pages a minute, and the cost analysis: ~$25 per million pages.