Tencent’s 4B vision model cuts out people and finds objects, with Apache weights

Tencent’s WeChat Vision team has published WeVisOne-VL-4B on Hugging Face, a vision-language model that handles object detection, counting, text reading, segmentation and image matting in one system. A 2B sibling appeared the same day.

The weights are released under Apache-2.0. The model is built on Alibaba’s Qwen3-VL and adds a small pixel decoder, so it can output cutout masks and alpha mattes without a separate segmentation model.

One model for boxes, masks and alpha mattes

The model card describes WeVisOne-VL, short for “WeChat Vision All-in-One”, as covering two levels of perception. At object level it does open-vocabulary detection, referring localization (finding the thing a sentence describes), counting, pointing and OCR with locations. At pixel level it does referring segmentation and promptable alpha matting, the soft-edged cutouts used for background removal.

Several design choices are spelled out. Boxes and points are written with dedicated coordinate tokens from <0> to <1000>, normalized to the image. The model answers None when the queried target is absent, and in dense tasks it emits an instance token, <OBJ>, after each predicted box.

The pixel work is done by a separate file, mask_decoder.pt, with about 9.6 million parameters in the 4B release (the 2B release’s decoder has about 9.5 million). It has a visual feature adapter, a box-guided query decoder with six deformable cross-attention blocks, and a head that produces full-resolution alpha mattes. The card lists the 4B checkpoint as 5B parameters in F32 Safetensors. The GitHub README speaks of 2B and 4B variants.

Tencent’s README for the Tencent/WeVisOne-VL repository gives results for the 4B model against the best baseline it lists, though it does not name those baselines in the summary we could read. Examples: LVIS detection F1 of 55.3 versus 50.7, Dense200 detection F1 of 71.3 versus 58.7, PixMo-Count accuracy of 89.6 versus 82.8, and ReasonSeg gIoU of 71.2 versus 65.0. On matting error, where lower is better, it reports 7.3 versus 15.4 on P3M-500-NP and 9.8 versus 22.4 on WeRefMat. Tencent says these used greedy decoding and at most 1,960 visual tokens per image.

Who can run it today, and how

Anyone can download the weights. The model card shows a Transformers example, vLLM and SGLang serve commands with OpenAI-compatible endpoints, and a Docker Model Runner line, docker model run hf.co/tencent/WeVisOne-VL-4B. A dedicated package from GitHub wraps the tasks, for example predict(image, task="matting", query="person"). The README lists Python 3.10, and devices including CUDA, NPU and CPU, plus a Gradio demo.

Tencent-authored code is under Apache License 2.0, while third-party components keep their own licenses. That matters for commercial users, who need to check each piece. Our explainer on open-weight models covers what a downloadable release does and does not give you.

What the release leaves unproven

The analysis here is about limits. The card itself carries no benchmark table and no paper link, and the figures above come from a README whose full results sit in a technical report we could not review. All of them are Tencent’s own claims, not independently reproduced. The repository showed 11 stars and a single commit when we read it, so there is no outside track record yet.

The base model sets a point of reference. According to its model card, Qwen3-VL-4B-Instruct is an Apache-2.0 model with 4B parameters and a native 256K context, and it advertises OCR in 32 languages and 2D grounding. WeVisOne-VL narrows the scope to single-turn image perception. Tencent says it has no explicit chain-of-thought reasoning, and its pixel decoder rejects video grids, so it is not a drop-in replacement for a general assistant.

Tencent also warns that fine alpha boundaries are sensitive to resolution, occlusion, complex transparency and prompt phrasing. Hair, glass and smoke are the usual hard cases for matting, so the headline error figures will need testing on other data. The card limits the intended use to research and says the model is not meant for autonomous high-impact decisions or for identifying individuals.

Two things would settle the claims: the technical report with named baselines and per-benchmark settings, and independent runs on matting sets that Tencent did not build. The README also notes that the Mask2Alpha matting annotator used in the data engine was open-sourced on a separate branch, and the fine-tuning code is not described as released. Developers who want to adapt the model to their own images will have to wait for that.