Tencent releases WeVisOne-VL-4B, a Qwen3-VL-based model for object detection, segmentation and alpha matting

Tencent has released WeVisOne-VL-4B, a unified vision-language model from WeChat Vision that covers both object-level and pixel-level perception. It is built on Qwen3-VL and adds a lightweight pixel decoder of 9.6M parameters (9,609,764 exactly) for segmentation and alpha matting. The model is listed on Hugging Face under tencent/WeVisOne-VL-4B.

What the model does

WeVisOne-VL (WeChat Vision All-in-One) unifies two groups of tasks through structured next-token prediction:

• Object-level perception: open-vocabulary object detection, referring localization, object counting, object pointing, and OCR.
• Pixel-level perception: referring segmentation and promptable alpha matting.

The Hugging Face page lists the model size as 5B params with F32 tensors, in Safetensors format. The model is tagged Image-Text-to-Text.

Architecture

The model uses dedicated coordinate tokens (<0> through <1000>) to represent normalized bounding boxes and points. It returns None when a queried target is absent. For dense tasks, it emits a special instance token, <OBJ> (token ID 152678), after each predicted box.

The attached mask_decoder.pt is a required core component. It combines three parts:

1. A Visual Feature Adapter operating over multi-level Qwen3-VL pre-merging ViT features.
2. A Box-Guided Query Decoder with 6 deformable cross-attention blocks, initialized from the <OBJ> hidden state and a Fourier-encoded box prior.
3. An Instance-Conditioned Dense Prediction Head using a FullCanvasAlphaCascadeHead, which refines predictions across S/4, S/2 and S/1 to output full-resolution continuous alpha mattes at stride 1.

Usage

Tencent provides an inference package that is installed from GitHub. The quick-start example loads the model with WeVisOneVL("tencent/WeVisOne-VL-4B", device="auto", dtype="bfloat16") and calls model.predict(image, task="matting", query="person", temperature=0.0). The result exposes a text response and a list of masks.

As of the page reading on 2026-10-23, no Inference Provider deploys the model. Pricing is not yet disclosed, and no benchmark scores or context window figures are given on the page.

Intended use and limitations

Tencent describes the release as intended for research in visual grounding, referring localization, pointing, counting, OCR localization, referring segmentation and image matting. It is not intended for autonomous high-impact decisions or for identifying individuals.

Tencent lists these limitations:

• Fine alpha boundaries remain sensitive to image resolution, severe occlusion, complex transparency and prompt phrasing.
• For inference efficiency, the model generates structured outputs directly, without explicit chain-of-thought reasoning.
• It currently targets single-turn image perception. Video grids are rejected by the pixel decoder.

License

Unless otherwise noted, Tencent-authored code in the repository is released under the Apache License 2.0. Third-party components keep their original licenses and copyright notices, and Tencent asks users to review the corresponding source files before use.

What this means

WeVisOne-VL-4B puts detection, pointing, counting, OCR localization, segmentation and matting behind one token-based interface instead of separate task-specific models. The pixel decoder is small at 9.6M parameters, but it is required: the language model alone does not produce masks or alpha mattes, so deployments must load mask_decoder.pt alongside the main weights. The single-turn, image-only scope and the lack of chain-of-thought mean the model targets direct perception outputs rather than multi-step visual reasoning. Without published benchmark figures on the page, its accuracy relative to other vision models cannot yet be assessed.