0.8B-Parameter OvisOCR3 Leads Five OCR Tasks at Once

A team of 15 authors has released OvisOCR3, a 0.8B-parameter model that handles five tasks with a single set of weights: document parsing, general OCR, key information extraction (KIE), captioning of figures in documents, and assembling tables that stretch across several pages. On OmniDocBench v1.6 it scores 96.74, and on OCRBench v2 it takes first place: 65.9 on the English split and 61.3 on the Chinese split. It is compared against general-purpose VLMs and narrowly specialized OCR models, and by the authors' reported results, the small model beats many of them.

Five teachers for one student

The main contribution is how the model was trained. The authors first ran continued pretraining on a mix of real and synthetic data. They then added domain-balanced mid-training. The tasks differ sharply in data volume and answer length. Parsing a full page into markup produces thousands of tokens, while extracting a couple of fields from a receipt takes only a few lines. Without balancing, the long domains would easily swamp the short ones, and this stage evens out each domain's contribution to training.

Next comes multi-teacher on-policy distillation. Five domain-specific teachers, each 4B in size, pass their skills to a general model that is five times smaller than any of them. On-policy distillation means the student trains on its own generations, which the teachers then evaluate. The final step is mixed-task reinforcement learning with per-domain rewards: tables, captions and field extraction each use different quality criteria.

Where OvisOCR3 comes out ahead

The authors compare the model on public and internal benchmarks across all five task families. Results according to their data:

• OmniDocBench v1.6 — overall score of 96.74.
• PureDocBench, Real5-OmniDocBench, MDPBench — the highest overall score among the compared models.
• OCRBench v2 — first place: 65.9 (English split) and 61.3 (Chinese split).
• OCRBench — higher than all baseline models.
• Four KIE benchmarks — higher than the baseline models on each.
• DocFigCapBench — the best localization F1 and caption quality close to Qwen3.8-Max.
• CrossTabBench — leads on Relation F1 for multi-page tables.
• MPDocBench-Parse — the best TEDS on truncated tables.

Figure captioning deserves a closer look. Here the 0.8B model nearly catches up with Qwen3.8-Max, and it beats every other model in the comparison on the accuracy of locating figures on the page. This task sits at the intersection of vision and language, where large general-purpose VLMs usually win.

Tables that don't fit on one page

Cross-page tables remain a weak spot for many OCR pipelines. A financial report, a specification or a statement often carries a table over to the next sheet with a repeated or missing header. A conventional parser sees two independent fragments, and the structure falls apart. Relation F1 on CrossTabBench and TEDS on the truncated tables of MPDocBench-Parse measure exactly this ability to restore the links between parts. Leading on both suggests the model understands the document as a whole, not page by page.

Who should take a look

One model instead of a set of specialized ones simplifies the architecture. You don't need separate services for parsing, KIE and image captioning, you don't have to reconcile their output formats, and you don't have to keep an on-call rotation for each. At 0.8B, the model promises cheap inference and the ability to run on modest hardware, although the authors do not publish cost comparisons, and you will have to check speed yourself.

Teams that build RAG over PDFs, process contracts, invoices and reports, or digitize Chinese and English documents get a candidate for a quick test. It is best to start with your hardest documents, the ones with multi-page tables and complex figures. Results on your own data will show how well the benchmark numbers carry over to production. The recipe itself, with domain teachers, distillation and RL with different rewards, also looks like a reproducible approach for other multi-domain tasks where several specialists need to fit into one compact model.