Post
80
WeVisOne-VL-4B on a Galaxy S26: Android developers can have their app box the objects a user names in a photo, with no detector to train.
litert-community/WeVisOne-VL-4B
By WeChat Vision at Tencent, built on Qwen3-VL: detection, pointing, counting and OCR.
In the clip, an offline test app runs it on the GPU: a 1967 NASA photo, Detect "person, hat", and 3 people and 2 hats boxed 7.3 s after the Ask tap.
To run it: load the 3.5 GB file with the litertlm-android 0.19.1 library, as in the clip; the card has the resize step and the task prompts.
litert-community/WeVisOne-VL-4B
By WeChat Vision at Tencent, built on Qwen3-VL: detection, pointing, counting and OCR.
In the clip, an offline test app runs it on the GPU: a 1967 NASA photo, Detect "person, hat", and 3 people and 2 hats boxed 7.3 s after the Ask tap.
To run it: load the 3.5 GB file with the litertlm-android 0.19.1 library, as in the clip; the card has the resize step and the task prompts.