MonkeyOCRv2 shrinks document OCR to 0.7B parameters and beats a 3B open-source model
量子位 · wechat · 2026-07-26
MonkeyOCRv2 pushes document OCR further into the small-model era: a 0.6B version scores 82.5 on MDPBench and a 0.7B version reaches 83.3, both ahead of the previous 3B open-source leader at 80.5.
- The team argues the gain comes less from sheer parameter count than from clearer division of labor: a stronger visual encoder captures characters, formulas, and layout more faithfully, while a smaller language model handles organization and understanding.
- Training combines image-to-text generation with pixel-level document reconstruction, so the vision tokens preserve enough evidence for strokes, symbols, and page structure instead of forcing the LLM to guess.
- They also open-sourced MonkeyDocv2, a 113M-document dataset spanning 17 languages, plus code, weights, demos, and deployment examples under Apache 2.0.
- The same encoder reportedly improves multiple downstream tasks, including OCR, formula recognition, document tampering detection, segmentation, parsing, and document understanding.
Related event: MonkeyOCRv2 Beats 3B Models with Sub-1B Parameters(2 posts)→
More from Infra
- Reddit asks whether RTX PRO 4500 makes sense beyond 24/7 low-power use — meikawaii · 2026-07-26
- Enterprises are shifting from token maxxing to budget maxxing — 0xsachi · 2026-07-26
- Receipts compare Google TPU, Gemini and xAI’s 10T training claims — imjustnewatai · 2026-07-26
- Predictor says $10,000 local-model desktops could become common next year — sull · 2026-07-26
- Three chokepoints dominate the global chip supply chain: Taiwan, ASML and Spruce Pine — sahilypatel · 2026-07-26
- A new report argues storage is becoming a core layer in agentic systems — BenBajarin · 2026-07-26