OvisOCR3: an 0.8B-parameter model that reads documents better than specialized systems

Why one small model instead of five specialized ones

The ATH-MaaS team has introduced OvisOCR3, a document recognition and analysis model weighing just 0.8 billion parameters that handles five different tasks at once and outperforms narrowly specialized systems on many metrics. The model is openly available on HuggingFace.

Parsing documents usually requires a whole toolkit: one model for printed text, another for tables, a third tuned for extracting fields from forms, a fourth for describing charts and images, and a fifth for stitching together tables split across PDF pages. Maintaining and serving five models is more expensive and more complicated than running one. But simply mixing all these tasks into a single training set doesn't work: a full page transcription runs to thousands of tokens of text, while extracting a single field from an invoice takes just a few words. If training loss is measured token by token, the long tasks drown out the short ones, and the model gets worse at precise data extraction.

The authors solved this with four-stage training. First, a model based on Qwen3.5-0.

8B was fine-tuned on 72 million examples — textbooks, reports, handwritten notes, and photos of documents taken on a phone. Then the dataset was rebuilt so each of the five tasks got roughly equal attention from the model (16–23% of 8.8 million examples per task), regardless of answer length.

Next came five "teachers" — separate 4-billion-parameter models, each an expert in its own narrow domain. They passed their knowledge to the final small model through what's called on-policy distillation: the model learns to correct its own mistakes under the teacher's supervision rather than simply copying the teacher's answers. The final stage was reinforcement learning, where the model received a separate reward for each of the five tasks, tuned to that task's specifics.

What the model can do in practice

OvisOCR3 covers five areas:

• Document parsing — reconstructs page structure: text, LaTeX formulas, HTML tables, reading order, and block layout on the image
• General text recognition (OCR) — ten different modes, from plain text reading to translating signage, recognizing formulas, parsing charts, and visual question answering with reasoning
• Key information extraction — pulls fields out of forms, invoices, and medical documents, either following a user-defined schema or independently identifying what matters in the document
• Document image captioning — locates photos, illustrations, charts, and flowcharts on a page and gives each a short text description (up to 100 words), building a relationship diagram for flowcharts
• Cross-page table reconstruction — if a table in a report is split by a page break, the model detects the split and merges the fragments back into one complete table

Benchmark results

On OmniDocBench v1.6, which measures accuracy on text, formula, and table recognition as well as reading-order correctness, the model scored 96.74 out of 100 — the best result among the systems tested. On three other document-parsing benchmarks (PureDocBench, Real5-OmniDocBench, MDPBench), it also came out ahead of every model it was compared against.

On OCRBench v2, a benchmark that tests text recognition under difficult conditions in two languages, OvisOCR3 took first place with 65.9 points on English and 61.3 on Chinese. It also beat its competitors on all four key information extraction (KIE) benchmarks and on the standard OCRBench.

For captioning images inside documents, the model delivered the best accuracy at pinpointing image boundaries, with caption quality close to that of Qwen3.8-Max — a noticeably larger model. On the task of merging tables split across pages, OvisOCR3 outperformed competitors on accuracy in reconstructing links between table fragments (CrossTabBench) and on the quality of reconstructing truncated tables (MPDocBench-Parse).

What this means for engineers

The practical takeaway: you don't need a zoo of five different models for five different document types — a single compact 0.8-billion-parameter model, trained the right way, can cover the entire range of tasks at a level matching specialists. That cuts infrastructure costs and simplifies deployment for teams that need to process mixed document streams: scans, photos, forms, and reports with tables and charts. The authors are candid that this result wouldn't be possible on a compact model without rebalancing the training data across tasks and without distillation from specialized teacher models — simply mixing datasets together isn't enough.