Skip to main content

Aliyun's Tiny 0.8B Model OvisOCR2 Tops Document Parsing Benchmarks

Aliyun has released OvisOCR2, an open-source document parsing model with just 0.8 billion parameters. Despite its small size, it scored 96.58 on the OmniDocBench v1.6 benchmark, ranking first and becoming the first end-to-end model to beat traditional pipeline approaches. This marks a breakthrough in document intelligence.

Image

For years, document parsing relied on a multi-step pipeline: first analyze layout, then recognize content. This approach had high maintenance costs, error accumulation, and complex deployment. OvisOCR2 changes the game by using a single model to directly output Markdown with text, formulas, tables, and visual areas in natural reading order.

Built on Qwen3.5-0.8B, the model combines real and synthetic data, and uses supervised fine-tuning (SFT), reinforcement learning (RL), online policy distillation (OPD), and model ensembling to unlock its full potential.

Key Points:

  • OvisOCR2 is a 0.8B parameter end-to-end document parsing model.
  • Achieved 96.58 on OmniDocBench v1.6, ranking first.
  • First end-to-end model to surpass traditional pipeline methods.
  • Uses SFT, RL, OPD, and model ensembling for performance.