Jina AI launches jina-ocr-v1: 25-language OCR scoring 91.14 on OmniDocBench
JinaAI_ · x · 2026-09-18
Jina AI released jina-ocr-v1, post-trained on DeepSeek-OCR with stronger multilingual support across 25 languages and optimization for budget GPUs like the NVIDIA L4.
Key results:
- 91.14 on OmniDocBench v1.6
- 83.4 on olmOCR-Bench
- Highest page throughput among competitors
The model is available on Hugging Face, alongside a tech report and blog post.
Related event: Jina AI Releases Open-Source OCR Model jina-ocr-v1(3 posts)→
More from Multimodal
- World Labs' Atlas reconstructs NVIDIA HQ in 3D from just 32 images, in real time — drfeifei · 2026-09-18
- fal made an open-source video model 35x faster — and Hollywood became its fastest-growing segment — a16z · 2026-09-18
- Pika relaunches as all-in-one AI creative platform aggregating Veo 3.1, Seedance, Nano Banana — aziz4ai · 2026-09-18
- First fully H3-generated Halloween kids music video drops, ditching Cling and WAN — R34vspec · 2026-09-18
- OpenAI's image update: fast Flare and precise Sunburst editing, 3 test images for $0.08 — paw_lean · 2026-09-18
- Epic god battle video made with GPT Image 2.5 Sunburst + Seedance 2.5 in Magnific — azed_ai · 2026-09-18