Tencent has released WeVisOne-VL, an open vision-language model, available in 2B and 4B parameter versions, according to the company’s WeChat Vision team. A single model handles detection, referring, counting, pointing, text recognition (OCR), segmentation, and fine matting (alpha matting). Pixel-level outputs are produced by a 9.6 million parameter decoder, without relying on a separate SAM model.
WeVisOne-VL, whose name stands for ‘WeChat Vision All-in-One’, is built on Qwen3-VL. It is designed for fine-grained localization of elements in an image rather than for free-form conversation. The user selects the desired output type: points, boxes, binary masks, or continuous alpha matte. When the requested object is absent from the image, the model responds ‘None’ instead of proposing a random box.
Operation and Training
Each task is treated as a structured prediction of the next token, i.e., the next unit of text or image the model generates. Box and point coordinates are expressed by tokens ranging from to . For pixel-related tasks, the model emits a token after each predicted box, whose hidden state serves as a query to the decoder. The latter includes a visual feature adapter, which reuses multi-level features from Qwen3-VL’s ViT, a box-guided query decoder composed of 6 deformable cross-attention blocks, and a dense prediction head that refines the result to a full-resolution mask or matte.
The model was trained using a knowledge graph-driven data engine, which annotated over 100 million images with approximately 3.6 billion boxes. Regions were proposed by SAM3, Rex-Omni, and an internal Tencent detector, then verified by Qwen3-VL. The training includes three stages of SFT (Supervised Fine-Tuning) and a final GRPO stage, on 178,000 examples filtered based on their uncertainty, as described in the associated research paper.
Announced Performance and Implementation
All following figures are provided by Tencent. For the 4B version, the model achieves an average F1 score of 71.3 on Dense200, dense detection, compared to 58.7 for LocateAnything-3B. On RefCOCOg test, it reaches 81.8 versus 77.6 for LocateAnything-3B. The rejection rate, which measures the ability to not propose a box for an absent object, reaches 96.4% on WeRef-Bench, an internal Tencent benchmark with 2,880 queries, compared to 73.0%. According to the source, the rejection rate is, however, 52.6% on PR-Bench, where many queries for absent targets still receive a box.
The weights for the 2B and 4B models are available on Hugging Face under the Apache 2.0 license. The 2B model shows 2.44 billion parameters, and the 4B model 4.83 billion. According to the source, inference runs on a single GPU, via Transformers or vLLM, in BF16. A community GGUF version handled object-level tasks on a 6GB RTX 2060 card. Tencent has also released Mask2Alpha, the matting annotation tool from its data engine. The fine-tuning code is yet to come.
