Tencent releases WeVisOne-VL-2B model on HuggingFace for image-text-to-text tasks

The model tencent/WeVisOne-VL-2B with the image-text-to-text task appeared on HuggingFace on October 22, 2026. This is a unified vision-language model from WeChat Vision, built on Qwen3-VL: it combines object-level perception — open-vocabulary object detection, referring localization, object counting, object pointing and OCR — with pixel-level tasks like referring segmentation and promptable alpha matting through structured next-token prediction and a lightweight pixel decoder.

What it means

The model contains 2.4 billion parameters, the largest weights file is model.safetensors at 9.1 GB, and the repository totals 9.1 GB. The weights are published in pt and safetensors formats, and it runs via transformers. License — apache-2.0. There are no quantized builds on the platform yet, so running it locally will require VRAM for the full weights. There is also a separate mandatory component, mask_decoder.pt, with 9,511,460 parameters — without it, pixel-level tasks do not work. You can follow the appearance of quantized builds and other news about running such models locally in the CPU3D news feed.