Tencent Open-Sources 2B Vision Model: Detection, Segmentation, and Matting in One Model

Tencent's WeChat Vision team announced WeVisOne-VL on X, and the code, weights, and technical report are already available on GitHub. Project URL: https://github.com/Tencent/WeVisOne-VL

The vision model industry has a habit: one model for detection, one for referring localization, one for segmentation, and another SAM for matting. The longer the pipeline, the more deployment feels like a kitchen filled with single-function small appliances. WeVisOne-VL aims to change this by folding object-level perception and pixel-level perception into a single vision-language model, available in 2B and 4B versions.

The tasks it supports are specific: open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation, and prompted alpha matting. Object-level tasks output boxes and points; pixel-level tasks output binary masks or continuous alpha mattes. According to the team, you no longer need to attach SAM separately for these tasks.

There are three key design pieces. The data engine uses a knowledge graph to discover concepts, multiple expert models to localize, and then a VLM to verify; it expands boxes into masks and alpha mattes, with annotations covering more than 100 million images. Tasks are unified into structured text, with coordinates represented by tokens from <0> to <1000>, and each box followed by an <OBJ> token. The pixel decoder has only 9.6M parameters; it takes the hidden states of the <OBJ> token, the predicted box, and multi-level Qwen3-VL ViT features to output full-resolution masks or alpha mattes.

Another practical point is rejection. When the target does not exist, the model returns None instead of forcing a box. In referring detection and referring segmentation, hallucination is more annoying than missed detection, and this design eliminates a lot of post-processing rules.

The performance numbers come from the team's technical report. WeVisOne-VL-4B achieves 55.3 F1 on LVIS detection, versus 50.7 for the strongest open baseline; 71.3 F1 on Dense200 dense detection, versus 58.7; 81.8 referring F1 on RefCOCOg, versus 77.6; a 96.4% rejection rate on WeRef-Bench, versus 73.0; 89.6% counting accuracy on PixMo-Count, versus 82.8; 52.0 OCR F1 on TotalText, versus 43.3; and 71.2 gIoU on ReasonSeg referring segmentation, versus 65.0. For matting tasks, P3M-500-NP achieves 7.3 MAD, versus 15.4, where lower MAD is better; WeRefMat referring matting achieves 9.8 MAD, versus 22.4. The reproduction entry point is evaluation/README.md in the repository; it is recommended to run it yourself before use.

Usage is not complicated. After installation, you can call the inference API directly; detection and matting share the same model loading logic:

from PIL import Image
from wevisone_vl import WeVisOneVL
from wevisone_vl.visualization import save_result

model = WeVisOneVL("tencent/WeVisOne-VL-4B", backend="transformers")
image = Image.open("demo/examples/images/livingroom.jpg").convert("RGB")
result = model.predict(image, task="detection", query=["sofa", "pillow", "coffee table", "lamp"])
save_result("outputs/detection", image, result)

image = Image.open("demo/examples/images/seabreeze.jpg").convert("RGB")
result = model.predict(image, task="matting", query="person")
save_result("outputs/matting", image, result)

The official release includes 2B and 4B online demos on Hugging Face, as well as a local Gradio launch option. The repository also open-sources Mask2Alpha, an annotation tool that converts binary masks into high-quality alpha mattes. The code uses Apache 2.0, third-party components retain their original licenses, and you should check the model card before using the weights.

For developers, the appeal of this project is a small model plus a lightweight pixel decoder that compresses detection, OCR, segmentation, and matting into a single inference chain, eliminating multi-model orchestration and SAM calls. For ordinary users, it is enough to understand it as 'one model that can find objects, count them, read text in images, and cut out people or objects.' Whether it can truly replace existing pipelines depends on latency, VRAM, and failure cases in real scenarios. At least judging from the form of the open-source release, Tencent has pushed the path of unified visual perception one step forward.