OvisOCR2 is a 0.8B open-weight document-parsing model from Alibaba's ATH-MaaS team, released on 2026-07-15. It takes a full page image and returns a Markdown file in natural reading order, covering body text, tables, math formulas, and figure regions.

Is OvisOCR2 open source and what license does it use?

Yes. OvisOCR2 ships under Apache-2.0 on Hugging Face at ATH-MaaS/OvisOCR2, with the training-and-inference code in the ATH-MaaS/Ovis GitHub repo. Weights are ~0.9B in Safetensors format and can be served with Transformers, vLLM, SGLang, or Alibaba's Docker image.

How does OvisOCR2 compare to pipeline OCR systems?

OvisOCR2 scores 96.58 on OmniDocBench v1.6, the first end-to-end model to top a leaderboard that had been dominated by pipeline systems chaining detector, layout, table, and formula tools. On PureDocBench it also leads with an Avg3 of 75.06, matching or beating far larger multimodal models.

How was OvisOCR2 trained?

OvisOCR2 post-trains Qwen3.5-0.8B with a hybrid dataset of real documents and HTML-derived synthetic pages, then applies supervised fine-tuning, reinforcement learning on a larger 4B teacher, on-policy distillation back to the 0.8B student, and model fusion. Rewards score text fidelity, formula accuracy, and table structure separately.

Can OvisOCR2 run on a laptop or on device?

OvisOCR2 is aimed at edge deployment: at 0.8B parameters it fits on a single consumer GPU and, with quantization, on high-end laptops or phones. That is the point of the release — an end-to-end parser small enough for local use that still beats server-side pipeline stacks on public benchmarks.

Alibaba · 2026-07-15 · major

OvisOCR2 — 0.8B Alibaba model tops OmniDocBench and beats pipeline OCR

OvisOCR2 is a 0.8B end-to-end document parser from Alibaba's ATH-MaaS team that turns page images into clean Markdown and scores 96.58 on OmniDocBench v1.6, the first end-to-end model to beat pipeline stacks on that leaderboard.

OvisOCR2 model card banner on Hugging Face — ATH-MaaS's 0.8B document parser

Alibaba's 0.8B end-to-end document parser sets state of the art on OmniDocBench v1.6.

Key specs

Parameters	0.8B
Omni doc bench v1.6	96.58
Pure doc bench avg3	75.06

Quick facts

Maker	Alibaba (ATH-MaaS team)
Parameters	0.8B
Base model	Qwen3.5-0.8B (post-trained)
License	Apache-2.0
Output	Markdown (text, tables, formulas, figures)
Runtimes	Transformers, vLLM, SGLang, Docker
OmniDocBench v1.6	96.58 (SOTA)

Benchmarks

OmniDocBench v1.6

OvisOCR2 (0.8B)		96.58

source ↗

What is it?

OvisOCR2 is a compact 0.8B open-weight document-parsing model from Alibaba's ATH-MaaS team. Given a page image, OvisOCR2 emits a single Markdown file in natural reading order that already includes body text, tables, formulas, and figure regions — no separate detector, layout, table, or formula pipeline required.

How does it work?

The model post-trains Qwen3.5-0.8B on a mix of real documents and HTML-derived synthetic pages. Training runs supervised fine-tuning, then reinforcement learning on a larger 4B teacher, then on-policy distillation back down to 0.8B, then model fusion. Rewards separately score text fidelity, formula accuracy, and table structure, so the tiny model learns to preserve each format instead of chasing one blended similarity score.

Why does it matter?

OvisOCR2 hits 96.58 on OmniDocBench v1.6, the first end-to-end model to top a leaderboard that had been dominated by pipeline systems chaining several specialists. Because the whole parser weighs 0.8B and ships Apache-2.0 with vLLM and SGLang support, teams doing document ingestion can now run a SOTA extractor on a single consumer GPU instead of stitching detector-plus-layout-plus-formula stacks together.

Who is it for?

teams building RAG, document AI, or on-device document extraction

Frequently asked questions

What is OvisOCR2?: OvisOCR2 is a 0.8B open-weight document-parsing model from Alibaba's ATH-MaaS team, released on 2026-07-15. It takes a full page image and returns a Markdown file in natural reading order, covering body text, tables, math formulas, and figure regions.
Is OvisOCR2 open source and what license does it use?: Yes. OvisOCR2 ships under Apache-2.0 on Hugging Face at ATH-MaaS/OvisOCR2, with the training-and-inference code in the ATH-MaaS/Ovis GitHub repo. Weights are ~0.9B in Safetensors format and can be served with Transformers, vLLM, SGLang, or Alibaba's Docker image.
How does OvisOCR2 compare to pipeline OCR systems?: OvisOCR2 scores 96.58 on OmniDocBench v1.6, the first end-to-end model to top a leaderboard that had been dominated by pipeline systems chaining detector, layout, table, and formula tools. On PureDocBench it also leads with an Avg3 of 75.06, matching or beating far larger multimodal models.
How was OvisOCR2 trained?: OvisOCR2 post-trains Qwen3.5-0.8B with a hybrid dataset of real documents and HTML-derived synthetic pages, then applies supervised fine-tuning, reinforcement learning on a larger 4B teacher, on-policy distillation back to the 0.8B student, and model fusion. Rewards score text fidelity, formula accuracy, and table structure separately.
Can OvisOCR2 run on a laptop or on device?: OvisOCR2 is aimed at edge deployment: at 0.8B parameters it fits on a single consumer GPU and, with quantization, on high-end laptops or phones. That is the point of the release — an end-to-end parser small enough for local use that still beats server-side pipeline stacks on public benchmarks.

Try it

huggingface.co/ATH-MaaS/OvisOCR2