AI
WeChat team's vision model handles detection, text reading and cut-outs in one
Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu, Zhipeng Huang, Chong Sun · Submitted 22 Oct 2026 · 239 upvotes yesterday
Preprint · not peer reviewedTrending
WeVisOne-VL is a vision-language model that does object detection, counting, text reading (OCR), segmentation and matting through next-token prediction, with a small pixel decoder for masks.
Key finding: The authors report leading results across over 30 benchmarks, while still refusing when the asked-for object is absent.
Why it matters: Today's vision chat models are unreliable at precise visual tasks; one model that can point, count and cut out objects would simplify many apps.