Flow - 4x faster than typing
Mini toolsOvisOCR2
By Alibaba
Model family: Qwen
OvisOCR2 is a 0.8B parameter end-to-end model built by post-training Qwen3.5-0.8B for page-level document parsing. Given a document page image, it outputs a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. It was trained using a data engine combining real-world and synthetic data with a multi-stage recipe integrating supervised fine-tuning, reinforcement learning, and online preference distillation. On OmniDocBench v1.6 it scores 96.58 overall, the highest among end-to-end models on that leaderboard. On PureDocBench it achieves an Avg3 score of 75.06. The model is compatible with vLLM for inference.
New
Multimodal
Overview
OvisOCR2 is a compact 0.8B parameter end-to-end model for page-level document parsing. Given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. Built by post-training Qwen3.5-0.8B with a multi-stage SFT, RL, and OPD training recipe.
About Alibaba
Chinese e-commerce and cloud leader behind Taobao, Tmall, and Alipay.
View Company Profile
Last updated: July 14, 2026
