Tencent open-sources WeVisOne-VL for all-in-one vision tasks

WeVisOne-VL is an open-source visual-language foundation model from Tencent's WeChat Vision Team that packs object detection, pointing, counting, OCR, and hair-level matting into a single lightweight checkpoint.

Traditionally, running a pipeline like this meant stitching together multiple heavy specialist models: an open-vocabulary detector for bounding boxes, an external segmenter like SAM for contours, and a dedicated matting network for soft transparencies. WeVisOne-VL collapses that entire pipeline into one unified system.

What it does

Feed it an image and a natural language prompt, and it handles the rest. You can ask it to detect furniture and receive bounding boxes, count objects with exact coordinates, or spot and read tricky text phrases like street signs.

Its standout capability is pixel-level perception. It generates continuous alpha mattes capable of capturing delicate hair strands without relying on SAM or other external heavyweights. On top of that, it features reliable rejection: if you query something absent from the image, the model cleanly outputs None instead of hallucinating annotations.

On the benchmark board, the 4B checkpoint consistently tops comparable 3B to 9B open models. It hits an F1 Mean of 71.3 on Dense200 dense detection, compared to 58.7 from strong baselines. On the WeRef-Bench rejection test, its rejection rate reaches 96.4%, avoiding the frequent false positives seen in other VLMs. Its matting MAD scores drop down to 7.3 on P3M-500-NP and 9.8 on WeRefMat referring tasks, where lower means cleaner edges.

Under the hood and how to run it

Architecturally, the model builds on Qwen3-VL and expresses all perception tasks as structured text tokens. The clever design choice is its pixel decoder: instead of adding a multi-hundred-million-parameter segmentation head, Tencent designed a tiny 9.6M-parameter decoder. It fuses token hidden states, bounding boxes, and multi-level ViT features to reconstruct full-resolution masks and alpha mattes with negligible compute overhead.

The project is released under the permissive Apache 2.0 license, permitting commercial adoption. Both 2B and 4B model weights are up on Hugging Face, alongside inference code, Cookbooks, benchmark evaluation scripts, and Mask2Alpha, the data engine annotator used to train the model.

In terms of hardware, the 4B model weighs around 19GB, while the 2B version is roughly half that size, making both easy to deploy on a single consumer GPU or local workstation. You can run the provided Gradio script locally right after cloning the GitHub repo, with built-in support for both Transformers and vLLM backends.

Why it matters

For developers building local image-editing tools, automated asset pipelines, or desktop vision agents, WeVisOne-VL is an easy recommendation to test right now. Its biggest achievement is architectural efficiency: by replacing the clumsy combination of a detector plus SAM plus a matting network with a sub-10M parameter decoder, it dramatically slashes VRAM footprints and operational latency.

Native target rejection is another huge practical win that prevents automated downstream workflows from breaking on absent objects. That said, as a 2B/4B model, its complex linguistic reasoning is ultimately bounded by its compact base LLM, so you should treat it as a sharp, highly specialized perception workhorse rather than a freeform reasoning engine.