MonkeyOCRv2 Beats 3B Models with Sub-1B Parameters

MonkeyOCRv2 pushes document OCR into the small model era, with its 0.7B version outperforming 3B open-source models. Additionally, controlled experiments reveal that swapping only the vision encoder results in a massive 13.2-point performance gap across eight benchmarks.

2026-07-26 ~ 2026-07-26 · 2 related posts

• MonkeyOCRv2 shows document visual encoders can decide 13.2 points of accuracy — 机器之心 · 2026-07-26
• MonkeyOCRv2 shrinks document OCR to 0.7B parameters and beats a 3B open-source model — 量子位 · 2026-07-26