OvisOCR3, a 0.8B model, covers five OCR task families
Original title: OvisOCR3 Technical Report
Weights are on Hugging Face; authors report 96.74 on OmniDocBench v1.6 and first place on OCRBench v2.
The OvisOCR3 technical report introduces a 0.8B model for document parsing, general OCR, key information extraction, document figure captioning, and cross-page table reconstruction. Training uses continued pretraining, domain-balanced mid-training, distillation from five 4B teachers, and mixed-task reinforcement learning. On the authors' comparisons it scores 96.74 on OmniDocBench v1.6, leads several named benches, and scores 65.9 and 61.3 on OCRBench v2 English and Chinese. Weights are at huggingface.co/ATH-MaaS/OvisOCR3.
Unconfirmed: Scores are the authors' comparisons on public and in-house benchmarks.
- Five domain-specific 4B teachers distill into the shared 0.8B model.
- OCRBench v2 scores are 65.9 English and 61.3 Chinese.
- It leads Relation F1 on CrossTabBench and truncated-table TEDS on MPDocBench-Parse.
- On DocFigCapBench it has the highest localization F1, near Qwen3.8-Max on captions.
- It also beats listed baselines on OCRBench and all four KIE benchmarks.
- Organizations
- ATH-MaaS
- AI models
- OvisOCR3Qwen3.8-Max
From the original article
1 min read at huggingface.co
We introduce OvisOCR3, a 0. 8B model for a broad range of OCR-related tasks. OvisOCR3 unifies document parsing, general OCR, key information extraction (KIE), document figure captioning, and cross-page table reconstruction in a single model.
By Yu Xia, Yinglun Li, Yuhui Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.