OCR models / Alibaba (Ovis)
OvisOCR3
Alibaba (Ovis) · released 2026-10-21 · data checked 2026-11-06
OvisOCR3 is a 850M-parameter OCR and document-parsing model from Alibaba (Ovis) (CN), released 2026-10-21 under the Apache-2.0 license. It scores 96.74 on OmniDocBench v1.6 (#5 of 33). It needs about 2.7 GB of memory at 16-bit.
Built on Qwen3.5 0.8B. Markdown, structured JSON with boxes, LaTeX, Mermaid for flowcharts and chart JSON.
OmniDocBench v1.6
96.74
#5 of 33 models
olmOCR-Bench
—
Not reported
Parameters
850M
OCR specialist
Memory to run
~2.7 GB
16-bit weights plus about 1 GB
Specifications
- Lab
- Alibaba (Ovis) (CN)
- Type
- OCR / document parsing specialist
- Total parameters
- 850M
- License
- Apache-2.0
- Memory to run
- ~2.7 GB at 16-bit
- Released
- 2026-10-21
Closest alternatives
- Xiaomi-OCR-0 (870M, 96.83 on OmniDocBench, Apache-2.0)
- TeleOCR (1.4B, 96.91 on OmniDocBench, Apache-2.0)
- Youtu-Parsing-Omni (5.3B, 96.96 on OmniDocBench, Youtu-Parsing license)
- OvisOCR2 (850M, 96.47 on OmniDocBench, Apache-2.0)
Sources
Something wrong here? Report it.