Alibaba's OvisOCR3: a 0.8B model tops OmniDocBench at 96.74 via multi-teacher distillation + RL
ATH-MaaS · hf · 2026-10-28
Alibaba's ATH-MaaS team released the OvisOCR3 technical report on Hugging Face: a compact 0.8B vision-language model unifying document parsing, general OCR, key information extraction, figure captioning, and cross-page table reconstruction.
Training recipe: continued pretraining on real and synthetic data, domain-balanced mid-training to balance supervision across tasks, multi-teacher on-policy distillation from five 4B domain-specific teachers, then mixed-task RL with domain-specific rewards.
Results: 96.74 overall on OmniDocBench v1.6, plus highest overall scores on PureDocBench, Real5-OmniDocBench and MDPBench, beating general-purpose VLMs and specialized OCR models.
More from Multimodal
- Seedance 2.5 hailed as taking AI filmmaking to another level — SimplyAnnisa · 2026-10-28
- Hands-on with Baidu's Unlimited OCR 2: whole docs in one pass, no page limit — Baidu_Inc · 2026-10-28
- Agentic video editing goes pro: senior editors 5x output, costs to collapse — dotey · 2026-10-28
- CodeCap-Bench: executable code beats natural language as video captions — Jiafu Tang · 2026-10-28
- Moving Atoms Engine turns real-world spaces and cities into high-resolution 4D models for sim — Scobleizer · 2026-10-28
- QuiverAI Launches Arrow 2 Vector Model on FLORA with Layered SVG Editing — andpoul · 2026-10-28