WeVisOne-VL-2B by tencent

WeVisOne-VL-2B by tencent

Reads an image with a text query and returns object boxes, points, counts, text positions and pixel masks, including alpha mattes. Suits researchers running a 2.4B model on their own computer.

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-2B is a vision-language model from the WeChat Vision team at Tencent. You give it an image and a text query, and it answers with object boxes, points, counts, text positions, segmentation masks or alpha mattes. It builds on Qwen3-VL and adds a small pixel decoder of 9.6M parameters for the mask work.

Who it suits: It suits researchers and engineers who work on visual grounding, counting, text localization, segmentation or image matting and want one model for all of them.

Industry | Who uses it | How they use it | Example
Computer vision research | Researchers | Test grounding, pointing and counting on their own image sets | Checking which box the model draws for a phrase
Photo and design tools | Developers | Cut a named subject out of a picture with an alpha matte | Separating a person from the background
Document processing | Engineers | Find where text sits in an image | Marking the region of a printed line
Data labeling | Annotation teams | Draft boxes and masks that a person then reviews | Pre-marking objects before manual checks

You can skip it if you need chat answers, video analysis or a way to identify people, since it handles single images and is not meant for that.

Where to use it:

• Python scripts that load the model through its inference package
• Windows, macOS or Linux with Python, for downloading the files
• Any Python environment where you can run the quick start script on an image

Use case | Who | What they get back
Referring segmentation | Researchers | A mask for the object a phrase describes
Alpha matting | Developers | A continuous alpha matte at full resolution
Counting and pointing | Analysts | Counts and points for the objects you name
Text localization | Engineers | Boxes that show where text sits in an image

How to set it up: You need Python and room for roughly 4.9 GB of weights in BF16.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• The wevisone_vl inference package from the project's GitHub page

1. Install huggingface_hub.
2. Download the model files to your computer.
3. Install the wevisone_vl inference package from the project's GitHub repository, following its own instructions.

pip install -U huggingface_hub
hf download tencent/WeVisOne-VL-2B

How to use it: Load the model, open an image, then call predict with a task and a query. The script below asks for a matte of a person and prints the response text and the number of masks returned.

from PIL import Image
from wevisone_vl import WeVisOneVL

model = WeVisOneVL("tencent/WeVisOne-VL-2B", 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 on a device with a chosen data type
predict | Runs a task on an image for a query and returns a response with masks

Example: You point the script at a photo called example.jpg and ask for a matte of a person. The call returns the model's text response and a list of masks, and the script prints the text and how many masks came back.

For newcomers and experts: If you are new to this, think of the model as a helper that finds things in a picture and outlines them. A first small try is the script above with a photo of one person, then reading how many masks it prints.

Under the hood, positions of boxes and points are written as dedicated coordinate tokens, and the model returns None when the target is absent. After each box in dense tasks it emits an instance token, and a 9.6M-parameter decoder turns that state and the box into masks. The trade-off is no step-by-step reasoning and single-turn image input only.

Image and text query go in
Qwen3-VL backbone writes boxes and points
Instance token marks each predicted box
Pixel decoder builds masks and alpha mattes
You read the response and masks

Good to know: The weights take about 4.9 GB in BF16, and about 1.2 GB at 4-bit as an estimate.

• Limits: fine alpha edges are sensitive to image resolution, heavy occlusion, complex transparency and how the prompt is worded
• License: Apache 2.0 for the Tencent-written code, and third-party parts keep their own licenses
• Cost to run: none named
• Data it reaches: the images you pass in and the downloaded model files
• Not for: autonomous high-impact decisions, identifying people, or video input

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 2.4B
• Weights: about 4.9 GB in BF16; about 1.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 15,085
• Likes: 85
• License: apache-2.0
• Quantized versions: 6
• Fine-tunes: 2
• 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-2B
• Files: https://huggingface.co/tencent/WeVisOne-VL-2B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/WeVisOne-VL-2B

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