Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
The 0.8B model scores 96.58 on OmniDocBench v1.6 and ships as open weights on HuggingFace, per the arXiv paper.
Why now: The paper and open weights landed together on arXiv and HuggingFace in July 2026.
OvisOCR2, an 0.8B parameter model, scored 96.58 overall on OmniDocBench v1.6, according to a paper posted to arXiv. It also posted a top Avg3 of 75.06 on PureDocBench.
That's a leaderboard multi-stage OCR pipelines have owned, and those pipelines charge real money per page. A model this small runs locally on a laptop or even a phone. No API bill, no per-page invoice.
The model is end-to-end, one pass in, Markdown out, covering text, formulas, tables, and visual elements. Weights are open on HuggingFace under a CC BY 4.0 license, so anyone can download and run the model without a subscription or an API key.
If OvisOCR2 holds up outside benchmark conditions, per-page document-ingestion pricing doesn't have much runway left. Watch whether the big OCR APIs cut prices or add a small-model tier in response.
The paper and weights landed together in July 2026, so the real test is whether outside teams can reproduce the OmniDocBench and PureDocBench numbers on their own documents, not just the benchmark.
Each link below shares sources, entities, or timing with this story.
HuggingFace released Claude Code / Shared entity: Markdown / Same source domain / Earlier coverage
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; reported by the same outlet (arxiv.org).
HuggingFace released Claude Code / Shared entity: Markdown / What happened next / Tension
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-08-11.
HuggingFace acquired GGML / Shared entity: HuggingFace / Earlier coverage
Linked by a graph relationship (HuggingFace acquired GGML); both cover HuggingFace; earlier HuggingFace coverage from 2026-02-22.
HuggingFace released Claude Code / Shared entity: Markdown / What happened next
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-08-30.
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-08-16.
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-07-30.
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-07-27.
Linked by a graph relationship (HuggingFace released Claude Code); both cover Markdown; picks up the Markdown thread on 2026-07-27.