Tencent’s WeChat Vision team has released WeVisOne-VL, an open vision language model available in 2B and 4B sizes. The new model handles detection, referring, counting, pointing, optical character recognition, segmentation, and alpha matting within a single architecture. Pixel outputs are generated from a 9.6 million-parameter decoder, eliminating the need for a separate Segment Anything Model.
Built on Qwen3-VL, the perception-focused model casts every task as structured next-token prediction and allows users to choose output granularity including points, boxes, binary masks, or continuous alpha mattes. Tencent trained the system using a knowledge-graph-driven data engine that annotated over 100 million images with roughly 3.6 billion boxes. According to self-reported tables, the model outperforms same-scale baselines across more than 30 benchmarks, such as achieving 71.3 F1 Mean on Dense200 dense detection.
Developers can access the model weights on Hugging Face and run the system via Transformers or vLLM backends on a single GPU. This unified approach matters because it allows developers to replace multiple specialized tools, such as an OCR engine, a detector, and a segmenter, with one streamlined model.