OvisOCR2-F32-GGUF

OvisOCR2, developed by the ATH-MaaS team, is a compact 0.8B vision-language model for end-to-end page-level document parsing, built by post-training Qwen3.5-0.8B to directly convert document images into structured Markdown while preserving the natural reading order. It accurately extracts plain text, complex tables (as HTML), mathematical formulas (as LaTeX), and visual regions, making it suitable for OCR, document understanding, and digitization workflows. Despite its small size, OvisOCR2 achieves state-of-the-art performance with an overall score of 96.58 on OmniDocBench v1.6, becoming the first end-to-end model to surpass traditional pipeline-based approaches on the benchmark, while also leading PureDocBench with an Avg3 score of 75.06. Trained using a combination of real-world and synthetic data with a multi-stage recipe incorporating Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and On-Policy Distillation (OPD), OvisOCR2 delivers high-quality document parsing with a lightweight deployment footprint and native support for efficient inference through vLLM.

Model Files

llama.cpp