0.8B-Parameter OvisOCR3 Outperforms OCR Models: 96.74 on OmniDocBench v1.6

Researchers at ATH-MaaS have introduced OvisOCR3, a compact 0.8B-parameter vision-language model that replaces a whole set of narrowly specialized OCR systems. A single set of weights parses documents, recognizes text in arbitrary images, extracts key fields from forms and receipts, describes charts and illustrations in documents, and stitches together tables that span several pages. On OmniDocBench v1.6 the model scored 96.74, and on OCRBench v2 it took first place. The weights and inference examples are published on Hugging Face in the ATH-MaaS/OvisOCR3 repository.

Five 4B-parameter teachers for one student

The central question of the project is how to fit five different skills into 0.8B parameters. The authors first continue training the model on a mix of real and synthetic data (continued pretraining). Next comes a domain-balanced mid-training stage, because the tasks differ sharply in data volume and answer length. A page parsed into Markdown and a short field extracted from a receipt are texts of different orders of magnitude, and without balancing, the long tasks would easily overwhelm the short ones.

Then comes multi-teacher on-policy distillation. Five domain teachers with 4B parameters each, one per task area, transfer their knowledge to the shared model. On-policy distillation means the student generates its own answers while the teachers evaluate and correct them. The pipeline ends with mixed-task reinforcement learning using domain-specific rewards: each type of task gets its own quality metric.

Where the model came out ahead

The authors compared OvisOCR3 with general-purpose VLMs and with specialized OCR models on public and internal benchmarks. Results across all five task families:

• OmniDocBench v1.6 (document parsing): overall score of 96.74.
• PureDocBench, Real5-OmniDocBench, MDPBench: the highest overall score among the compared models.
• OCRBench v2: first place, with 65.9 on the English split and 61.3 on the Chinese split.
• OCRBench: the model outperforms the baseline models used for comparison.
• Four key information extraction (KIE) benchmarks: best results on all four.
• DocFigCapBench (captions for document illustrations): best localization F1; caption quality is close to Qwen3.8-Max.
• CrossTabBench (tables spanning multiple pages): leads on Relation F1.
• MPDocBench-Parse: best TEDS for truncated tables.

First place on OCRBench v2 matters in its own right: the benchmark tests not just text reading but also different ways of working with text in two languages. The score on the Chinese split is lower than on the English one, but the model leads on both.

Tables that don't fit on one page

Cross-page table assembly deserves separate attention. Financial reports, specifications, and registries regularly carry rows over to the next sheet, and a typical page-by-page parser breaks the structure: the header stays on one page while the data lands on another. OvisOCR3 restores the links between fragments, and this is where it leads on Relation F1 and on TEDS for truncated tables. For tasks like loading documents into RAG or automating accounting, this is a pain point that used to be patched with workarounds.

The model arrives amid a wave of similar work. Hugging Face semantic search surfaces PolyOCR-Venus, Xiaomi-OCR-0, Jina-OCR-v1, and a paper on lightweight VLMs for OCR with structured JSON alongside it. The market is clearly moving toward small unified models instead of chains of a detector, a recognizer, and post-processing.

Who should try it

At 0.8B parameters, the model can run on modest hardware: a single consumer GPU, in a private environment without sending documents to an external API. That is of interest to teams that process scans of contracts, invoices, forms, and reports running to thousands of pages. One engine instead of five simplifies deployment and maintenance.

The limitations are also clear: the comparisons were run by the authors themselves, and some of the benchmarks are internal. So before changing a production pipeline, it is worth running the model on your own documents, especially those with rare fonts, poor scans, and multi-page tables. The authors themselves are asking the community for feedback, so the first independent measurements will appear quickly.