Alibaba Cloud Open-Sources OvisOCR2, Hits 96.58 on Benchmark

Alibaba Cloud has open-sourced OvisOCR2, a document parsing model that scored 96.58 on the OmniDocBench v1.6 test, becoming the first end-to-end model to beat traditional methods.

Document parsing is key for AI models to handle PDFs and scans. Old methods use multiple models, which are costly and error-prone. OvisOCR2 works in one step: input an image, get Markdown text with formulas and tables.

It's based on Qwen3.5-0.8B and trained with synthetic data. The project is open-source under Apache 2.0 and works with vLLM, fitting into the Qwen ecosystem.