OvisOCR2 is a Qwen3.5 0.8B fine-tune by Alibaba that is topping the benchmarks (for it's size) on multiple OCR tasks.
Context
78,222
tokens
OvisOCR2 is a fine-tuned version of Qwen3.5 0.8B by Alibaba that tops the benchmarks (for its size) on multiple OCR tasks. It is a compact end-to-end page-level document parsing model (with a size of only 0.8 billion parameters). Given a document page image, OvisOCR2 can generate Markdown representations that follow natural reading order, covering text, formulas, tables, and visual regions.
OvisOCR2 was developed following the completion of Qwen3.5-0.8B training. It employs a carefully designed data engine that combines real-world data with synthetic data, along with a multi-stage training scheme that integrates SFT (Supervised Fine-Tuning), RL (Reinforcement Learning), and OPD (Optimal Parameter Distribution). The model achieves powerful document parsing performance while maintaining a small deployment resource footprint.
What documents were tested with OvisOCR2?
The model was tested on invoices, receipts, and financial documents to demonstrate OCR performance on realistic document types.
How is OvisOCR2 loaded and run in this review?
The model is loaded and run using llama.cpp, a tool that enables local execution of language models. This allows users to run the model on their own systems.
