OvisOCR2 Tops Document Parsing Leaderboard
Littlepharaoh · reddit · 2026-07-15
The author tested OvisOCR2 (0.8B), an end-to-end document parsing VLM. Built on Qwen3.5-0.8B with continued pre-training and an Apache 2.0 license, it runs on vLLM 0.22.1. The model scored 96.58 on OmniDocBench v1.6, becoming the first end-to-end model to top the leaderboard, surpassing traditional "detection + crop + OCR + reassembly" pipelines.
End-to-end approaches avoid the single point of failure common in pipeline OCR, where missed regions during layout detection cause permanent text loss. Deployment is also simplified to a single model and serving stack.
In comparative tests using 827 real scanned medical documents, OvisOCR2 outperformed the GLM-OCR pipeline in overall word-level F1, showing significant improvements in tail percentiles where pipeline failures often stemmed from layout detector misses. The author also detailed throughput tests on a single RTX 5090, DPI trade-offs, fixes for repetitive output loops, and deployment specifics like mm-processor-cache and ninja.
Related event: OvisOCR2 Tops Document Parsing Leaderboard(2 posts)→
More from Multimodal
- Gen games don't need to be realtime — turn-based interactivity works, devs argue — djcows · 2026-10-27
- Short film 'Odissey of Aeons' made with MiniMax H3 video generation — Musikalandia · 2026-10-27
- Remiqora Video: open-source local studio turns one idea into 4K chained AI videos — inikolax · 2026-10-27
- Single HTML/JS file generates creative art and copy in-browser with zero LLM calls — Afinetheorem · 2026-10-27
- $5k kale-img-1 fine-tuning contest won with MOSS-Image base, SDXL trails far behind — ostrisai · 2026-10-27
- HN discussion: How far image generation has come since August 2022 — hunterbrooks · 2026-10-27