OvisOCR2: End-to-End Document Parsing Model
Shiyin Lu · hf · 2026-07-16
This is a technical report for 0.8B document parsing model OvisOCR2, designed to convert single-page document images directly into Markdown arranged in natural reading order, covering text, formulas, tables, and visual regions.
The authors built a data engine featuring:
- Filtered real-world document annotations
- Synthetic pages rendered from the same HTML source to ensure consistency between the image and the Markdown target
The training pipeline includes:
- supervised fine-tuning
- Reinforcement learning on a 4B branch with a multi-component reward design
- on-policy distillation back to the 0.8B model
- model fusion
As a result, OvisOCR2 achieved an overall score of 96.58 on OmniDocBench v1.6, securing the top spot at the time, and also scored the highest Avg3 of 75.06 on PureDocBench. Evaluations on custom long-tail and difficult scenario benchmarks also show leading results, indicating strong generalization and robustness beyond public leaderboards.
Related event: OvisOCR2 Tops Document Parsing Leaderboard(2 posts)→
More from Multimodal
- ComfyUI Krea2 Prompt Styler Node Revived in a Fork With a Massive Style Library — Winter_unmuted · 2026-10-27
- Single HTML file generates retro poster designs in-browser with zero LLM calls — Afinetheorem · 2026-10-27
- A Redditor Used Copilot to Imagine and Archive Their Own '80s/'90s Winter Streetwear Concepts — Valuable-Sky2134 · 2026-10-27
- Transkun open-source piano transcription model nails MIDI, sustain events included — alexanderchen · 2026-10-27
- fal adds FLUX 3 Image Klein LoRA training: 10 reference images, 3 LoRAs per request — OdinLovis · 2026-10-27
- Guy turns his girlfriend's selfie into a playable AI game ad about popping pimples — CurieuxExplorer · 2026-10-27