OCR systems often stop at transcription, while general vision-language models can be costly for document work. This model targets the gap with a compact checkpoint that handles both page structure and questions about what the page says. Its reported results suggest that OCR-centric understanding can complement document parsing rather than compete with it.
Key Capabilities
- Structured page parsing: Produces reading-order output with tables and formulas represented in structured formats, which helps preserve more than plain text during document conversion.
- Document understanding: Answers OCR-focused visual questions and extracts fields into a requested JSON schema, so the same model can support both reading and downstream data capture.
- Small model, broad task coverage: The 0.8B model reports an average of 83.2 across five OCR-VQA benchmarks, compared with 80.9 for Qwen3.5-2B in the model card’s comparison. Its reported parsing scores include 96.83 on OmniDocBench v1.6 and 95.24 on Real5-OmniDocBench.
- OCR-focused training pipeline: Training combines text-to-location anchoring, continued pretraining on an approximately 170-million-sample corpus, and mixed-task reinforcement learning. The data pipeline uses expert consensus and rendered-output checks to catch structural errors.
Who It’s For
A good fit for teams processing forms, reports, tables, scanned pages, or document images that need both extraction and question answering from a single model. Look elsewhere if you need a general-purpose vision model, or if you require independently verified results across your own documents: the benchmark figures are reported by the project, and domain-specific accuracy should be tested before deployment.
