Alibaba's $800M Document Model Takes the Lead, with One Model Outperforming the Entire Pipeline for the First Time

According to DongCha Beating monitoring, Alibaba Cloud has open-sourced the document parsing model OvisOCR2. It is trained based on Qwen3.5-0.8B and can directly convert a document page into Markdown, restoring text, formulas, tables, and reading order at once.

In the past, tasks like this usually required multiple steps such as layout analysis, content recognition, and result splicing. OvisOCR2 accomplishes this in a single step with an end-to-end model, making deployment simpler and reducing errors caused by multi-stage handoffs.

On OmniDocBench v1.6, OvisOCR2 achieved a comprehensive score of 96.58, surpassing the pipeline model PaddleOCR-VL-1.6 at 96.33, becoming the first end-to-end model to top the leaderboard. The model is released under the Apache 2.0 license, and the weights are available on Hugging Face and Model Zoo.