WeVisOne-VL-4B

WeVisOne-VL-4B

Finds, counts, points to, reads and segments objects in images, and makes alpha mattes, with one 4B vision-language model from Tencent. For researchers working on visual grounding and image matting.

Made by tencent on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: WeVisOne-VL is a vision-language model from WeChat Vision at Tencent, built on Qwen3-VL. It handles object-level tasks such as open-vocabulary detection, referring localization, counting, pointing and OCR. It also handles pixel-level tasks, namely referring segmentation and promptable alpha matting, using a small pixel decoder of 9.6M parameters.

Who it suits: It suits researchers and developers who work on visual grounding, segmentation and matting and want one model for many perception tasks.

Industry | Who uses it | How they use it | Example
Research | Vision researchers | Study unified object and pixel perception | Compare detection and matting in one model
Software | Developers of image tools | Add object finding or matting to an app | Cut a person out of a photo with a matte
Content | Image editors | Isolate a subject by name | Ask for the person in a picture

People who need video, multi-turn chat or reasoning chains can skip it, since the card limits it to single-turn image perception.

Where to use it:

• Python, using the inference package the card points to on GitHub
• Machines that can run a 4B model, with device and dtype set automatically in the card's script
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Object finding | Researchers | Boxes or points for the target you name, or None when it is absent
Referring segmentation | Developers | Masks for the object you describe
Alpha matting | Image editors | A full-resolution alpha matte for the subject

How to set it up: You need these first:

• Python and the huggingface_hub package
• The WeVisOne-VL inference package from the card's GitHub project
• Pillow, which the script uses to open images

1. Install the inference package that the card points to on GitHub.
2. Download the model files to your computer. They take about 9.7 GB in BF16.
3. Run the quick start script with your own image.

pip install -U huggingface_hub
hf download tencent/WeVisOne-VL-4B

How to use it: Load the model, open an image, and call predict with a task and a query. The card's script asks for a matte of a person with the temperature set to 0.0. It prints the response text and the number of masks.

from PIL import Image
from wevisone_vl import WeVisOneVL

model = WeVisOneVL("tencent/WeVisOne-VL-4B", device="auto", dtype="bfloat16")
image = Image.open("example.jpg").convert("RGB")
result = model.predict(image, task="matting", query="person", temperature=0.0)
print(result.response, len(result.masks))

Tool | What it does
WeVisOneVL | Loads the model with the device and dtype you choose
model.predict | Runs a task such as matting on an image with a text query
result.response | Holds the text the model returns
result.masks | Holds the masks, which the script counts

Example: You open a photo and ask for a matte of the person in it. The script prints the model's response and how many masks came back. The card says fine alpha edges can suffer with low resolution, heavy occlusion or complex transparency.

For newcomers and experts: This model looks at a picture and finds things you name, then can outline them down to the pixel. A first small step is to run the quick start script on one photo with a clear subject.

The model writes boxes and points as coordinate tokens on a normalized grid, and it returns None when the queried target is missing. After each predicted box it emits an instance token for dense tasks. The mask_decoder.pt file is required and holds 9,609,764 parameters. It has a feature adapter, a query decoder guided by boxes, and a dense head that sharpens the matte.

You give an image and a query
The model predicts boxes or points as tokens
The pixel decoder builds masks or mattes
You read the response and masks

Good to know:

• Size: 4.8B parameters per Hugging Face, about 9.7 GB in BF16 and about 2.4 GB at 4-bit, as an estimate
• Limits: it gives structured outputs without explicit reasoning, and the pixel decoder rejects video grids
• Versions: Hugging Face lists 4 quantized and 1 fine-tuned version made by other people
• License: apache-2.0, with third-party components keeping their own licenses
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face and reads the images you pass in
• Not for: autonomous high-impact decisions or identifying individuals, which the card says it is not intended for

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.8B
• Weights: about 9.7 GB in BF16; about 2.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 18,908
• Likes: 122
• License: apache-2.0
• Quantized versions: 4
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_vl, image-text-to-text, vision-language, object-detection, referring-expression-comprehension, object-counting, point-grounding, referring-expression-segmentation, image-matting, conversational

Links

• Model: https://huggingface.co/tencent/WeVisOne-VL-4B
• Files: https://huggingface.co/tencent/WeVisOne-VL-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/WeVisOne-VL-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.