OvisOCR3: open-source 0.8B OCR model adds structured JSON output and cross-page table reconstruction
Sad_External6106 · reddit · 2026-10-21
OvisOCR3 is out, built on Qwen3.5-0.8B, Apache 2.0 licensed, with vLLM inference examples. Beyond OvisOCR2's Markdown parsing it adds:
- Structured JSON output with bounding boxes, element categories, and content, plus a layout-only mode
- Key information extraction via specified fields or a JSON template
- Figure descriptions and Mermaid output for flowcharts
- Cross-page table reconstruction: parse pages first, then merge table fragments across adjacent pages with the same model
Reported scores: 96.74 on OmniDocBench v1.6, 88.80 on MDPBench (up from 69.00), and 82.31 on PureDocBench (up from 75.10). The poster cautions these are team-reported numbers, and asks for community comparisons with GLM-OCR, PaddleOCR-VL, and MinerU, plus VRAM and speed data on messy scans and long financial tables.
More from Models
- ChatGPT web can actually listen to audio, but the same model can't via API or Codex — Possible_Statement84 · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- TLA+ expert corrects Grok: TLC only exhausts finite state spaces, not all behaviors — tianyin_xu · 2026-10-23
- OpenAI blocks manual reset with 13% usage left: 'no reset needed right now' — AccomplishedSugar490 · 2026-10-23
- Meta's 816B Muse Spark Open scores 39, losing to 180B Qwen3.8-Flash-Next at 40 — BTA_Labs · 2026-10-23
- Gemini refuses to generate image in Docs, blames document size limit — viksit · 2026-10-23