What is OvisOCR2?
OvisOCR2 is a 0.8-billion-parameter document parser from ATH-MaaS, a team inside Alibaba Group, released in July 2026 under Apache 2.0 with a technical report on 15 July. It takes one page image and writes the whole page as Markdown in reading order, with formulas as LaTeX, tables as HTML and figures as tagged bounding boxes. There is no separate layout model in front of it.
It is a post-trained Qwen3.5-0.8B. The team trained with supervised fine-tuning, ran reinforcement learning on a larger 4B branch, distilled that branch back into the 0.8B model and then averaged several variants. Real-document labels came from PaddleOCR-VL-1.5 and MinerU2.5-Pro, converted to one Markdown format.
On 21 October 2026 the same team released OvisOCR3, also 0.8B on the same base, and OvisOCR2's model card now points to it as the newer version.
Is OvisOCR2 worth it?
OvisOCR2's headline score holds up outside its makers' lab, which is rarer than it should be in this category. The team reports 96.58 on OmniDocBench v1.6; OpenDataLab's own leaderboard, which runs the models itself, lists 96.47. That places it third on the leaderboard behind OXR-1.0 at 97.03 and TeleOCR at 96.91, and just ahead of PaddleOCR-VL-1.6 at 96.34 and GLM-OCR at 95.22, with fewer parameters than either of those.
The design argument is that a single pass beats a layout model plus a recogniser. The report's in-house test of more than 1,000 pages supports it: pipeline parsers dropped 13 to 17% of tables on the complex-table subset, against 7.96% for OvisOCR2, and its overall score of 85.54 beat 82.88 for PaddleOCR-VL-1.6. That benchmark is Alibaba's own and unpublished, and the model learned from PaddleOCR-VL-1.5 and MinerU2.5-Pro labels, so part of the gain is cleaner versions of its teachers' output.
The tooling is thin. The card offers one vLLM recipe pinned to version 0.22.1 and nothing for Transformers, SGLang or Ollama; no throughput figure, hosted API or fine-tuning guide is published. The sample code also ships a function to trim repeated text at the end of long outputs, a sign that the model can loop on dense pages, and the card asks users to check results by hand in critical work.
The bigger problem is OvisOCR3. It keeps the 0.8B Qwen3.5 base and the Apache 2.0 licence, and its card reports 96.74 on OmniDocBench v1.6, 82.31 on PureDocBench against OvisOCR2's 75.10, and 88.80 on MDPBench against 69.00, plus field extraction and cross-page tables. Those are self-reported and not yet on the public leaderboard, but a team that already runs OvisOCR2 under vLLM has little reason not to test the drop-in successor.
Our take: OvisOCR2 is the safer of the pair for the next few months only because its score has been checked by someone else; once OvisOCR3 appears on the OpenDataLab leaderboard, OvisOCR2 becomes a reference point rather than a choice.
Where it earns its place
An ML engineer at a fintech building a RAG index over scanned annual reports
Turn report pages with dense tables into Markdown without a layout-detection step
Serve the model with the card's vLLM recipe, render each page between 448 and 2,880 pixels a side, and keep the HTML tables for chunking. Alibaba's own test shows fewer missed tables than pipeline parsers, which matters most for financial statements.
A platform team at a university library with a single mid-range GPU
Digitise papers with formulas at high volume on hardware they already own
At about 1.7 GB in BF16 the model fits alongside other services on one card. Formulas come back as LaTeX; budget time to measure pages per hour yourself, since no throughput figure is published.
Before you commit
- Setup
- pip install vllm==0.22.1 and pillow, then load ATH-MaaS/OvisOCR2 with the card's fixed parsing prompt; thinking must be disabled in the chat template.
- Integrations
- vLLM only in the official instructions; community GGUF conversions exist on Hugging Face. Output is Markdown with HTML tables, LaTeX formulas and image tags whose coordinates are scaled to 0 to 1,000.
- Data and privacy
- Apache 2.0 for commercial use and fine-tuning. Self-hosted only, so documents stay on your hardware; the Hugging Face demo should not see confidential pages.
- Cost at scale
- Weights are free and about 1.7 GB in BF16, so one consumer or data-centre GPU serves it. No pages-per-second figure is published, so capacity planning needs your own measurement.
- Limits
- One page per call, output capped at 16,384 tokens in the sample; repetition on long pages has to be trimmed; the card asks for manual checks in critical applications.
What’s good
- 96.47 on OmniDocBench v1.6, independently measured by OpenDataLab
- About 0.85B parameters, roughly 1.7 GB in BF16
- One pass per page; no separate layout model to deploy
- Apache 2.0, with community GGUF builds
What’s not
- Superseded by OvisOCR3 within three months
- Only a vLLM recipe; no hosted API or throughput figures
- Sample code has to trim repetition loops on long outputs
- In-house benchmark and handwriting results are unpublished data
Best for
Teams that want a small, single-pass page parser whose accuracy has been confirmed on the public OmniDocBench leaderboard.
OvisOCR2 alternatives
OvisOCR2: common questions
- Is OvisOCR2 worth it?
- OvisOCR2 proved a 0.8B end-to-end model can match the best pipeline parsers, and three months later its own successor made it the second choice. We rate it 4.1 out of 5.
- How much does OvisOCR2 cost?
- Free under Apache 2.0. No hosted API; a free demo runs on Hugging Face Spaces.
- What are the downsides of OvisOCR2?
- Superseded by OvisOCR3 within three months; only a vLLM recipe; no hosted API or throughput figures; sample code has to trim repetition loops on long outputs; in-house benchmark and handwriting results are unpublished data.
- Who is OvisOCR2 best for?
- Teams that want a small, single-pass page parser whose accuracy has been confirmed on the public OmniDocBench leaderboard.
- How much does OvisOCR2 cost for a team or in production?
- Weights are free and about 1.7 GB in BF16, so one consumer or data-centre GPU serves it. No pages-per-second figure is published, so capacity planning needs your own measurement.
- Is OvisOCR2 safe for company data?
- Apache 2.0 for commercial use and fine-tuning. Self-hosted only, so documents stay on your hardware; the Hugging Face demo should not see confidential pages.
- What are the best OvisOCR2 alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, GLM-OCR, Xiaomi-OCR-0 and MiniCPM-OCR 1.0.
Answers checked Nov 2, 2026.
Make OvisOCR2? Show this rating on your site
