WeVisOne-VL

WeVisOne-VL (WeChat Vision all-in-One) is a vision-language model from Tencent that combines object-level and pixel-level perception in a single model. It handles object detection, referring localization, counting, pointing, OCR, segmentation and alpha matting, and returns "None" when the queried object is absent rather than hallucinating one.

It is built on Qwen3-VL. Coordinates are emitted as dedicated tokens (<0>…<1000>), so a box takes four tokens and a point two. A 9.6M-parameter pixel decoder is attached through a special <OBJ> token and reuses the VLM's own multi-level ViT features, with no SAM or external segmentation backbone. It consists of a visual feature adapter, a box-guided query decoder and an instance-conditioned dense head that produces masks and full-resolution alpha mattes.

Training used a three-stage curriculum of supervised fine-tuning followed by mixed-task reinforcement learning on roughly 100M images, with a data engine that combines localization experts, VLM verification and SAM2 for mask annotation. The authors report leading results on 30+ benchmarks. According to a YouTube review, the model is released under Apache 2.0 in 4B and 2B variants. The official page fetched did not link weights or code directly, so those details are unverified here.