Alibaba Open-Sources 0.8B Document Parsing Model OvisOCR2, Sets New Benchmark

July 24, 2026 – Alibaba has announced the open-sourcing of OvisOCR2, a document parsing model that has achieved a significant breakthrough in the field of artificial intelligence. The model has secured a comprehensive score of 96.58 on the OmniDocBench v1.6, setting a new record and becoming the first end-to-end model to surpass the long-standing "pipeline method" in document parsing.

End-to-End Breakthrough

Document parsing is a critical infrastructure for large language model (LLM) applications, essential for converting unstructured documents like PDFs and scanned images into structured text for use in corporate knowledge bases, RAG (Retrieval-Augmented Generation) systems, and intelligent Q&A.

Traditionally, the industry relied on "pipeline methods," which typically involve two separate models: a layout analysis model to identify document structure and a content recognition model to extract text, formulas, and tables. While established, this approach suffers from high maintenance costs, error accumulation, and deployment complexity.

OvisOCR2 adopts an end-to-end technical route. Instead of chaining multiple models, the new architecture takes a single document image as input and generates Markdown-formatted text in a natural reading order. This single-step process covers text, formulas, tables, and visual regions, marking a significant shift in the technology landscape.

Small Parameters, High Capability

Despite its compact size, OvisOCR2 demonstrates "small parameter, big ability." It is derived from the Qwen3.5-0.8B model through post-training. To achieve this performance, the team constructed a hybrid data engine that combines real-world documents with synthetic data designed to specifically address complex scenarios such as interwoven tables and formulas, multi-column layouts, and long output requirements.

The training process involved a multi-stage approach, integrating Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), Online Policy Distillation (OPD), and model fusion to fully unlock the potential of the compact model.

Ecosystem Integration

OvisOCR2 is released under the Apache 2.0 license, making it freely available for developers. The model is compatible with mainstream inference frameworks like vLLM and integrates seamlessly with the Qwen3.5 reasoning ecosystem. This allows developers to incorporate the model into their workflows without the need for extensive additional adaptation.