Tencent’s WeChat Vision team added the WeVisOne-VL-4B model to Hugging Face. It builds on Qwen3‑VL and can do open‑vocabulary object detection, counting, OCR, referring segmentation and alpha‑matte extraction in one step. The model includes a lightweight 9.6 M‑parameter pixel decoder and uses special coordinate tokens to return bounding boxes and masks. The release is aimed at research use only; it is not meant for autonomous high‑impact decisions.
Why it matters
Researchers can now experiment with a single model for many vision‑language tasks without stitching together separate systems.
