OvisOCR3: 0.8B Open-Weights Model Unifies Document Parsing, OCR, KIE and Cross-Page Tables
The Alibaba-backed Apache 2.0 checkpoint packs five document tasks into one small model, with adoption peaking this week; all benchmark figures remain vendor-reported.
OvisOCR3, a 0.8-billion-parameter multimodal document model built on Qwen3.5-0.8B, is drawing a wave of community attention this week as repositories and papers pages update around it. The weights were published under Apache 2.0 by ATH-MaaS on 21 October, with a technical report on arXiv, and a community GGUF build of roughly 812 MB makes it runnable on modest local hardware.
The model unifies five document tasks in a single checkpoint: document parsing to JSON, Markdown or HTML with LaTeX; general OCR with bounding boxes and visual question answering; key information extraction driven by JSON templates; document figure captioning including Mermaid output for flowcharts; and cross-page table reconstruction, predicting when a table has no continuation. Training ran through four stages ending in distillation from five specialist 4B teacher models and reinforcement learning with per-domain rewards.
Vendor-reported benchmarks place it at 96.74 on OmniDocBench v1.6 and 65.9 on OCRBench v2 English, ahead of larger models such as Qwen3.5-4B. The authors caution nothing has been independently audited, and English-Chinese dominance in training data means performance on other languages must be tested rather than assumed.