Tencent releases WeVisOne-VL-2B and 4B vision-language models for object and pixel perception
Article timeline
The articles in this story. Times are ET.
ModelScopeOfficialPickTencent WeChat Vision releases open-weight WeVisOne-VL-4B perception modelAIModelScope announces WeVisOne-VL-4B, an open-weight visual perception model from Tencent's WeChat Vision team, built on Qwen3-VL. The team says one model handles seven perception tasks, including detection, counting, pointing, segmentation and alpha matting, and reports 71.3 on Dense200, 89.6 on PixMo-Count and 71.2 on ReasonSeg. It also reports a 96.4% rejection rate on WeRef-Bench and is licensed under Apache-2.0.
- Tencent · new models on Hugging FaceOfficialTencent releases WeVisOne-VL-4B, a unified vision-language model for object and pixel perception
AITencent's WeChat Vision team releases WeVisOne-VL-4B, a vision-language model built on Qwen3-VL that handles object-level tasks such as open-vocabulary detection, referring localization, counting, pointing, and OCR, plus pixel-level referring segmentation and alpha matting. The model adds a 9.6M-parameter pixel decoder and is intended for research use under the Apache License 2.0 for Tencent-authored code.
- Tencent · new models on Hugging FaceOfficialTencent releases WeVisOne-VL-2B, a unified vision-language model for object and pixel perception
AITencent's WeChat Vision team releases WeVisOne-VL-2B, a vision-language model built on Qwen3-VL that unifies object-level tasks such as detection, counting, and OCR with pixel-level referring segmentation and alpha matting. A lightweight 9.5M-parameter pixel decoder handles the pixel tasks, and the model is available on Hugging Face under the Apache License 2.0 for research use. The model generates structured outputs without explicit chain-of-thought reasoning, and fine alpha boundaries remain sensitive to resolution, occlusion, and prompt phrasing.