What is OvisOCR3?
OvisOCR3 is a 0.8-billion-parameter OCR and document model from ATH-MaaS, a team inside Alibaba Group, released on 21 October 2026 under Apache 2.0, a day after its technical report. Like OvisOCR2 it is a post-trained Qwen3.5-0.8B that reads a page image in one pass, with no separate layout model in front of it. OvisOCR2's model card now points to it as the newer version.
What changed is scope. OvisOCR2 did one job, turning a page into Markdown. OvisOCR3 adds structured JSON parsing with a bounding box and category for every region, layout-only detection, ten general OCR prompts (from plain text and word boxes to formulas, chart data, translation and visual question answering), three modes of key-information extraction, figure captions with Mermaid code for flowcharts, and a two-page mode that stitches tables split across a page break. Each task is chosen by a prompt file shipped with the weights.
Training changed to match. The first stage used about 72 million examples, 57% of the sampling budget on document parsing and 29% on general OCR. Five 4B teachers, one per task family, were then distilled into the 0.8B student, followed by reinforcement learning across all five tasks at once.
Is OvisOCR3 worth it?
OvisOCR3 is the better model than OvisOCR2 on every number its makers publish, but none of those numbers has been checked by anyone else. The card reports 96.74 on OmniDocBench v1.6 against OvisOCR2's self-reported 96.58, a gap of 0.16 points, barely more than the 0.11 OpenDataLab knocked off OvisOCR2's figure when it ran the model itself. OpenDataLab's leaderboard, last updated on 11 September, does not list OvisOCR3, and on that one benchmark Alibaba's own table puts Xiaomi-OCR-0 slightly ahead at 96.83.
The larger gains are on harder pages. On PureDocBench, which tests the same pages clean, digitally degraded and re-photographed, the score rises from 75.10 to 82.31 and the drop from clean to recaptured pages nearly halves, from 15.04 points to 7.84. On MDPBench, 2,720 images in 17 language categories, it rises from 69.00 to 88.80, and the report says the gap between Latin and non-Latin scripts shrinks from 23.7 points to 1.5. These are public benchmarks, but Alibaba ran OvisOCR3 on them itself.
The new tasks are where the evidence is thinnest. Field extraction scores 90.50 on OCRBench's KIE subset and 88.69 on CC-OCR, the latter only 0.26 ahead of Qwen3.5-OCR. Cross-page tables reach 97.07 relation F1 against 59.14 for MinerU2.5-Pro, but on CrossTabBench, an in-house set of 890 page pairs; on the public MPDocBench-Parse the lead is 92.04 to 91.87. Figure captioning is scored on another in-house set of 182 pages whose reference captions were drafted by a vision model. AI Weekly notes that the report publishes no inference latency against the 4B teachers and does not separate what the distillation and reinforcement-learning stages each contributed.
The deployment story has not moved since OvisOCR2. The card gives one vLLM recipe pinned to 0.22.1 and no throughput figure. Markdown output still needs the repeated-tail cleanup function, extraction JSON is prompted rather than enforced, so it needs validating, and the card asks for generated HTML and Mermaid to be sanitised before rendering. Community builds arrived fast: mradermacher's GGUF conversion has 7,007 downloads, more than the 4,972 for the original weights.
Our take: for anyone already running OvisOCR2, this is a weights swap with more prompts, not a migration, and the extraction and cross-page modes may let one model replace two or three. Teams choosing a parser for the first time on accuracy alone are better served by GLM-OCR or PaddleOCR-VL-1.6 until OvisOCR3's numbers appear on a table Alibaba does not control.
Where it earns its place
An accounts-payable engineer at a mid-sized distributor receiving supplier invoices in several languages
Pull invoice number, date, seller and line items into JSON without a second model
Send each page with the schema-based extraction prompt and a JSON template, as the card shows for invoices, then validate the JSON before posting it. Alibaba reports 77.73 on its in-house KIE set against 75.25 for GLM-OCR; check accuracy on your own suppliers first.
A data team at an investment research firm parsing annual reports
Rebuild financial tables that break across page boundaries into single HTML tables
Parse each page with the structured prompt, then run the card's document_crosspage.py over adjacent pages, which merges pairs from the end backwards. Accuracy depends on the first-pass parse and correct page order, so keep the raw outputs for review.
A digitisation lead at a European archive working from phone photographs of records in non-Latin scripts
Turn photographed, non-Latin pages into searchable Markdown
Run the Markdown parsing prompt under vLLM on one GPU. The MDPBench gains are the strongest reason to prefer it over OvisOCR2 here, though they are Alibaba's own runs and no pages-per-hour figure is published.
Before you commit
- Setup
- pip install vllm==0.22.1 and pillow, download the repository with its prompts folder, and load a prompt file per task; thinking must be disabled in the chat template, and requests take one or two images.
- Integrations
- vLLM only in the official instructions; community GGUF, MLX and ONNX int4 conversions exist on Hugging Face. Outputs are Markdown, JSON, HTML tables, LaTeX and Mermaid, with boxes on a 0 to 1,000 grid.
- Data and privacy
- Apache 2.0 for commercial use and fine-tuning. Self-hosted, so documents stay on your hardware; the Hugging Face demo should not see confidential pages.
- Cost at scale
- Weights are free and the safetensors file is about 1.7 GB, so one GPU serves it. No throughput or latency figure is published, so capacity planning needs your own measurement.
- Limits
- Sample output is capped at 16,384 tokens; repeated tails must be trimmed; KIE JSON and HTML need validating; captions are English only, up to 100 words; the card warns of errors on small or blurred text, dense layouts and long tables.
What’s good
- Five task families in one 0.8B checkpoint, about 1.7 GB in BF16
- MDPBench up from 69.00 to 88.80 over OvisOCR2, per Alibaba
- Structured JSON with region boxes, plus layout-only mode
- Apache 2.0, with GGUF, MLX and ONNX community builds
What’s not
- Not yet on the OpenDataLab leaderboard; every score is self-reported
- Cross-page and captioning wins rest on in-house benchmarks
- Still only a vLLM recipe, with no throughput or latency figures
- Markdown output still needs repetition cleanup; JSON is not enforced
Best for
Teams that want parsing, field extraction and cross-page tables from one small self-hosted model, and can validate it on their own documents.
OvisOCR3 alternatives
OvisOCR3: common questions
- Is OvisOCR3 worth it?
- OvisOCR3 does five document jobs in the footprint OvisOCR2 used for one, and the only thing holding it back is that nobody outside Alibaba has scored it yet. We rate it 4.2 out of 5.
- How much does OvisOCR3 cost?
- Free under Apache 2.0. No hosted API; a free demo runs on Hugging Face Spaces.
- What are the downsides of OvisOCR3?
- Not yet on the OpenDataLab leaderboard; every score is self-reported; cross-page and captioning wins rest on in-house benchmarks; still only a vLLM recipe, with no throughput or latency figures; markdown output still needs repetition cleanup; JSON is not enforced.
- Who is OvisOCR3 best for?
- Teams that want parsing, field extraction and cross-page tables from one small self-hosted model, and can validate it on their own documents.
- How much does OvisOCR3 cost for a team or in production?
- Weights are free and the safetensors file is about 1.7 GB, so one GPU serves it. No throughput or latency figure is published, so capacity planning needs your own measurement.
- Is OvisOCR3 safe for company data?
- Apache 2.0 for commercial use and fine-tuning. Self-hosted, so documents stay on your hardware; the Hugging Face demo should not see confidential pages.
- What are the best OvisOCR3 alternatives?
- The closest alternatives we have reviewed are OvisOCR2, GLM-OCR, PaddleOCR-VL-1.6, Unlimited-OCR 2, LightOnOCR-3-4B and MiniCPM-OCR 1.0.
Answers checked Nov 2, 2026.
Make OvisOCR3? Show this rating on your site
