How to Run WeVisOne-VL-4B Locally: Tencent's All-in-One Vision Model

WeVisOne-VL-4B is a 4.83 billion parameter vision-language model from Tencent's WeChat Vision team, built on Qwen3-VL, released under Apache 2.0. One model handles object detection, referring localization, counting, pointing, OCR, segmentation and promptable alpha matting, and it runs locally through a pip-installable package with either a transformers or vLLM backend.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts, licenses, hardware requirements and install commands against the primary source before recommending a setup path.

A new top-20 entrant on Hugging Face's image-text-to-text list

WeVisOne-VL-4B entered the top 20 of Hugging Face's image-text-to-text trending list roughly a week after Tencent published it on October 22, 2026. At time of writing the model page shows 87 likes and 7,306 downloads, and the GitHub repository backing it has 461 stars. The most-liked Space linked to the model is fastgpu/vram-and-gpu-cost-calculator (15 likes), a general GPU cost estimator rather than a demo of the model itself. For an actual hands-on demo, Tencent hosts its own 4B Space you can try before installing anything.

Specs at a glance

What it actually does

Most vision-language models are one trick each: a detector, or a segmenter, or an OCR model, bolted together when a project needs more than one. WeVisOne-VL-4B is built to be the whole toolbox in one checkpoint. It handles open-vocabulary object detection, referring expression localization, object counting, pointing, OCR, referring segmentation, and promptable alpha matting, all through the same predict() call with a different task argument. Ask it to find a category from a list and it returns boxes. Ask it to find "the red mug on the left" and it does referring localization instead, or returns None if nothing in the image matches.

That None behavior is a deliberate design choice, not a gap. Most vision models will hallucinate a bounding box for a query that doesn't match anything in the image rather than admit it found nothing. WeVisOne-VL-4B is trained to reject absent-target queries explicitly, and the model card backs that up with a benchmark built specifically to measure it.

Architecture: a pixel decoder bolted onto a VLM

The model pairs a Qwen3-VL vision-language backbone with a separate 9.6 million parameter pixel decoder (9,609,764 parameters exactly, shipped as mask_decoder.pt) that turns language-model outputs into pixel-level masks and alpha mattes. Object-level tasks use coordinate tokens written as <0> through <1000> to represent normalized box and point positions directly in the token stream, and a dedicated <OBJ> token (ID 152678) marks where a pixel-level prediction should attach. The pixel decoder itself combines a Visual Feature Adapter over multi-level Qwen3-VL features, a Box-Guided Query Decoder with six deformable cross-attention blocks, and a cascade head that refines predictions across three resolution stages to produce a full-resolution alpha matte.

a lightweight 9.6M-parameter pixel decoder, enabling localization at any user-specified granularity from points and boxes to binary masks and continuous alpha mattes, while reliably rejecting absent-target queries

How it benchmarks against other open models

Tencent's own GitHub repository publishes head-to-head numbers against Qwen3-VL, Qwen3.5, Rex-Omni and LocateAnything, the strongest open baseline in each case. WeVisOne-VL-4B leads on every listed benchmark.

The rejection rate gap is the widest one on the sheet: 96.4 percent versus a 73.0 percent best baseline on WeRef-Bench, a benchmark built specifically to test whether a model will correctly say no target found instead of guessing. The GitHub repository also publishes a smaller 2B variant. The 4B model leads on most benchmarks, but the 2B model edges it out on a handful of pointing and box-matting metrics, so it's worth checking both detail tables before picking a size on benchmark performance alone.

Install the inference package

Run detection and matting

The package ships with sample images and a single predict() call that takes a task name and a query. This example runs open-vocabulary detection against a list of categories, then runs promptable alpha matting on a different image using a plain text query instead of a box.

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

model = WeVisOneVL("tencent/WeVisOne-VL-4B", backend="transformers")  # or backend="vllm"
# model = WeVisOneVL("tencent/WeVisOne-VL-2B", backend="transformers")  # Use 2B instead

# Object-level: open-vocabulary detection
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)

# Pixel-level: promptable alpha matting
image = Image.open("demo/examples/images/seabreeze.jpg").convert("RGB")
result = model.predict(image, task="matting", query="person")
save_result("outputs/matting", image, result)

task selects which of the seven supported modes runs: detection, referring, point, count_box or count_point, ocr or ocr_grounding, segmentation, or matting. device accepts "auto", "cuda:0", "npu:0", or "cpu", and the usual decoding arguments like temperature and max_new_tokens are also exposed if you need to tune generation behavior for a specific task.

Launch the Gradio demo

If you'd rather test the model through a browser UI than write a script, the repository includes a Gradio demo you can run locally against either model size and either backend.

License and intended use

Tencent-authored code and weights ship under Apache 2.0, so commercial use, modification, and redistribution are permitted without asking permission. The model card scopes intended use to research in visual grounding, referring localization, pointing, counting, OCR localization, referring segmentation, and image matting, and explicitly says it isn't intended for autonomous high-impact decisions or for identifying individuals.

Known limitations

No VRAM figure is published, so you're estimating hardware fit rather than reading it off the card. The model generates structured outputs directly without chain-of-thought reasoning, which keeps inference fast but means it won't reason through an ambiguous query the way a slower, reasoning-oriented VLM might. It's also built for single-turn image perception only: the pixel decoder rejects video grids outright. Fine alpha matte boundaries stay sensitive to image resolution, heavy occlusion, complex transparency, and how a referring expression is phrased.

FAQ

What is WeVisOne-VL-4B?

A 4.83 billion parameter vision-language model from Tencent's WeChat Vision team, built on Qwen3-VL. It unifies object detection, referring localization, counting, pointing, OCR, segmentation, and alpha matting in one checkpoint rather than requiring separate specialist models for each task.

Is WeVisOne-VL-4B free for commercial use?

Yes. It ships under Apache 2.0, which allows commercial use, modification, and redistribution without needing permission from Tencent.

How much VRAM does WeVisOne-VL-4B need?

Neither the Hugging Face model card nor the GitHub README publishes a minimum figure. A smaller 2B variant is also available if the 4B model doesn't fit your hardware.

What's the difference between the 2B and 4B versions?

Both share the same architecture and supported tasks. WeVisOne-VL-4B leads on most published benchmarks, including an 81.8 versus 78.2 F1 Mean edge over the 2B model on RefCOCOg referring, though the 2B model wins on a few pointing and box-matting metrics. The 2B model is also the lighter option for memory-constrained hardware.

Where to go from here

• Browse other local vision-language models: the local AI model directory lets you compare specs side by side.
• New to running vision-language models locally? the how-to guide index has setup guides for other local models that cover similar install and inference workflows.