Today, Alibaba Cloud officially open-sourced and released the document parsing model OvisOCR2.
With a comprehensive score of 96.58, the model has set a new record on the OmniDocBench v1.6 leaderboard, becoming the first end-to-end model to surpass pipeline methods and claim the top spot. This marks a significant breakthrough in the technical approach within the field of document parsing.
Figure: OvisOCR2's overall and categorical performance on OmniDocBench v1.6.
End-to-End Model Surpasses Pipeline Methods for the First Time
Document parsing is a crucial infrastructure for the real-world deployment of large models. Scenarios such as enterprise knowledge bases, RAG (Retrieval-Augmented Generation) retrieval, and intelligent Q&A all require the conversion of unstructured documents like PDFs and scanned files into structured text.
Previously, this domain was dominated by "pipeline methods," typically consisting of a layout analysis model and a content recognition model. The former identifies the document's layout, while the latter recognizes text, formulas, tables, and other content, which are then spliced together for the final output. Although this approach is mature, it has inherent bottlenecks, including high maintenance costs, error accumulation, and deployment complexity.
OvisOCR2 adopts an end-to-end technical approach: given a document image as input, the model outputs a Markdown representation following natural reading order in a single generation pass, covering text, formulas, tables, and visual regions. Work that previously required multiple models to collaborate is now accomplished in one step by a single model.
Topping the OmniDocBench v1.6 with a score of 96.58, it proves for the first time that an end-to-end model can outperform pipeline methods.
Small Parameters, High Capability: Achieving SOTA with 0.8B
OvisOCR2 is derived from the post-training of Qwen3.5-0.8B. The team constructed a data engine system that complements real documents with synthetic ones. The synthetic documents are specifically designed to supplement complex scenarios such as interleaved tables and formulas, multi-column layouts, and long outputs. The training protocol integrates a four-stage process: Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), Online Policy Distillation (OPD), and Model Merging—forming a multi-stage, progressive training pipeline that fully unlocks the potential of the compact model.
Figure: OvisOCR2 Data Engine System
Project Open-Sourced and Released, Seamlessly Integrating into the Qwen Ecosystem
OvisOCR2 is open-sourced under the Apache 2.0 license. It is compatible with mainstream inference frameworks like vLLM and connects directly with the Qwen3.5 inference ecosystem, allowing developers to integrate it without the need for additional adaptation.
Model Access:
Hugging Face: https://huggingface.co/ATH-MaaS/OvisOCR2
ModelScope: https://modelscope.cn/models/ATH-MaaS/OvisOCR2
Technical Report: https://arxiv.org/abs/2607.13639
/ END /