Tencent releases WeVisOne-VL-2B, a 2B-parameter model for object detection, segmentation and alpha matting

Tencent has released WeVisOne-VL-2B, a 2B-parameter unified vision-language model that covers both object-level and pixel-level perception in a single system. Developed by WeChat Vision, Tencent Inc., the model is built on Qwen3-VL and published on Hugging Face under the name tencent/WeVisOne-VL-2B. WeVisOne stands for WeChat Vision All-in-One.

What the model does

According to the model card, WeVisOne-VL unifies two groups of tasks:

• 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 model handles these through structured next-token prediction plus a lightweight pixel decoder with 9.5M parameters. The attached mask_decoder.pt file (9,511,460 parameters) is a required core component, not an optional add-on.

Architecture details

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 pixel decoder combines three parts:

1. A Visual Feature Adapter that operates 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 with a FullCanvasAlphaCascadeHead, which refines predictions across S/4, S/2 and S/1 to output full-resolution continuous alpha mattes at stride 1.

Usage and licensing

The model is loaded through an inference package installed from GitHub. The Quick Start shows a WeVisOneVL class that takes the model ID, a device setting and a dtype of bfloat16, and a predict call that accepts an image, a task (for example matting), a text query and a temperature. The result exposes a text response and a set of masks. The weights are distributed as Safetensors with F32 tensors.

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

Intended use and limitations

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

Tencent lists several 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.

Benchmark scores and a context window figure are not stated on the model page. Pricing is not yet disclosed, and the model page lists no Inference Provider deployment.

What this means

WeVisOne-VL-2B packages detection, counting, OCR localization, segmentation and matting into one model with a small added decoder, rather than requiring separate specialist models for each task. The design choice is to keep the language model backbone from Qwen3-VL and add pixel output through a 9.5M-parameter module keyed off the <OBJ> token. Developers working on grounding or matting can evaluate one checkpoint for several tasks, but without published benchmark numbers on the model page, comparisons against task-specific models will need independent testing. The single-turn, image-only scope and the lack of chain-of-thought reasoning set clear boundaries for where the model fits.