Tencent WeVisDoc and Jina AI's jina-ocr-v1 Launch on the Same Day
On September 18, Hugging Face welcomed two document OCR models on the same day: Tencent's WeVisDoc and Jina AI's jina-ocr-v1, both scoring highly on OmniDocBench; jina-ocr-v1 trended on the HF hot list the next day.
Confirmed
- Tencent released tencent/WeVisDoc in 2B and 4B versions under Apache 2.0; the 4B model scored 95.38 on OmniDocBench, which @NielsRogge says tops the OCR leaderboard
- Jina AI released jina-ocr-v1, an image-text-to-text multimodal model (based on deepseekvlv2) with 3.4B total and only 570M active parameters and speculative decoding; non-commercial license
- jina-ocr-v1 is post-trained on DeepSeek-OCR, with support for 25 languages and inference optimized for low-cost GPUs such as the NVIDIA L4
- jina-ocr-v1 scored 91.14 on OmniDocBench v1.6; it handles PDFs, scans, tables, charts and invoices and outputs clean Markdown
- On September 19, jina-ocr-v1 climbed the Hugging Face trending list
Why it matters
- With both models launching the same day, users can weigh parameter size, license (Apache 2.0 vs non-commercial) and language coverage when choosing
- jina-ocr-v1's low active-parameter design plus speculative decoding dramatically lowers deployment barriers for budget GPU environments
- WeVisDoc's 95.38 shows small models can achieve top OCR accuracy
2026-09-18 ~ 2026-09-19 · 6 related posts
Primary sources
- [source] Jina AI launches jina-ocr-v1, a 570M-active-param visual document parser — JinaAI_ · 2026-09-18
- Jina AI launches jina-ocr-v1: 25-language OCR scoring 91.14 on OmniDocBench — JinaAI_ · 2026-09-18
- [source] jina-ocr-v1 post-trained on DeepSeek-OCR, scores 91.14 on OmniDocBench — JinaAI_ · 2026-09-18
- [source] Two new OCR models land on Hugging Face: WeVisDoc-4B scores 95.38 on OmniDocBench — NielsRogge · 2026-09-18
- Jina AI's jina-ocr-v1 document understanding model trends on Hugging Face — jinaai · 2026-09-19
1 near-duplicate retellings: NielsRogge