OvisOCR2 parses full documents with a small local model
OvisOCR2 is a document-reading model with about 800 million , built on Qwen3.5-0.8B. It turns a full document page directly into structured Markdown while preserving text, tables, formulas, and reading order.
Its model card reports scores of 96.58 on v1.6 and 75.06 on . The are available under the , and the model supports vLLM.
These results are strong for its size, but they come from the published model information and still need independent testing on real documents. Direct comparisons with GLM-OCR, PaddleOCR-VL, MinerU, and Chandra are also still needed.
Key points
- The model has about 800 million and processes complete document pages.
- It converts text, tables, formulas, and reading order into structured Markdown.
- Reported scores are 96.58 on v1.6 and 75.06 on .
- Its weights use the , and vLLM is supported.
- Independent real-world tests and direct comparisons with competing models are still needed.