# Document-Understanding Foundation Models

Labs increasingly pretrain large multilingual vision models specialized for document parsing and OCR, pushing structured document understanding toward a general foundation-model approach; expect recurring open releases posting benchmark gains on multilingual document parsing.

- Conviction: 35 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-07-15T00:00:00.000Z
- Last updated: 2026-07-15T14:01:30.699Z
- Canonical: https://polylog.news/ai/trends/document-ai-foundation-models
- Publisher: Polylog

## Recent evidence

- [confirms] MonkeyOCRv2 Pretrains a Document Vision Model on 113 Million Images Across 17 Languages (2026-07-15): MonkeyOCRv2 pretrains a document vision model on 113 million images across 17 languages using a dual image-to-text and pixel-reconstruction objective, reporting a 2.8-point absolute gain on the MDPBench multilingual parsing benchmark over the prior best open model.