WeChat team open-sources 4B vision model: counting, character recognition, fine-grained matting.

Beating AI News Flash: Tencent WeChat Vision Team has open-sourced WeVisOne-VL, releasing 2B and 4B versions. It can perform object detection, counting, OCR text recognition, object localization by description, image segmentation, and fine matting all within a single model. For example, if asked to find the person wearing red clothes in a photo, the AI can mark their location and then extract the person's outline. If the person does not exist in the image at all, the model can also answer None, reducing the occurrence of hallucinated object markings.

Traditional solutions typically require a relay of detection models, SAM segmentation models, and matting models to complete the task. WeVisOne-VL is based on Qwen3-VL, encoding object coordinates into special Tokens, and then through a pixel decoder with only 9.6 million parameters, generates segmentation masks and alpha mattes with transparency, handling details such as hair. The entire process requires no additional calls to SAM.

In tests published by Tencent, the 4B version achieved an F1 score of 71.3 on Dense200 dense object detection, surpassing the strongest listed competitor's 58.7; on PixMo-Count, counting accuracy reached 89.6%, compared to 82.8%. In the team's self-built WeRef-Bench, when faced with non-existent targets, the rejection rate for false positives reached 96.4%.