Tencent's WeChat Vision team released WeVisOne-VL, an open vision-language model available in 2B (2.44B params) and 4B (4.83B params) sizes on Hugging Face, built on Qwen3-VL to unify detection, referring, counting, pointing, OCR, segmentation, and alpha matting. The model casts all tasks as structured next-token prediction using coordinate tokens (0–1000) for boxes and points, emitting "None" for absent targets. For pixel-level outputs, it employs a 9.6M-parameter decoder comprising a Visual Feature Adapter reusing Qwen3-VL ViT features, a 6-block Box-Guided Query Decoder with deformable cross-attention seeded by an token and box coordinates, and a Dense Prediction Head refining outputs at S/4, S/2, and S/1 resolutions. Training leveraged a knowledge-graph-driven engine annotating over 100M images with ~3.6B boxes across three SFT stages—format alignment (90M examples), scaled visual tokens (1,960) and box capacity (500), and pixel decoder training plus matting specialization (130M samples)—followed by GRPO on 178K uncertainty-filtered examples. Self-reported benchmarks (greedy decoding, ≤1,960 visual tokens) show the 4B model achieving 71.3 F1 on Dense200 detection vs. 58.7 for LocateAnything-3B, 81.8 F1 on RefCOCOg referring vs. 77.6, 96.4% rejection on WeRef-Bench, 89.6% exact-match counting on PixMo-Count, 52.0 F1 on TotalText OCR surpassing PP-OCRv5's 25.7, 71.2 gIoU on ReasonSeg segmentation vs. LISA-13B's 65.0, and 7.3 MAD on P3M-500-NP matting vs. SAM2-Matting's 15.4. The 2B model trails on most tasks but scores 6.6 MAD on box-prompted matting. Models run on a single GPU via Transformers or vLLM in BF16; a community GGUF enabled object tasks on a 6GB RTX 2060, though GGUF lacks mask support. Weights are Apache 2.0 licensed.