The Pulse
Tencent Releases WeVisOne-VL in 2B and 4B Sizes
Tencent’s WeChat Vision team released WeVisOne-VL, a vision-language model that combines object detection, OCR, segmentation and image matting. Its authors report a 9.8 mean absolute difference score on a referring-matting benchmark, compar
AI.info Team · Written with AI · fact-checked ·
WeVisOne-VL-4B scores 9.8 on a referring-image-matting test where a larger Qwen3.5-9B paired with SAM2-Matting scores 22.4, according to results released by Tencent’s WeChat Vision team. Lower scores are better on the benchmark’s mean absolute difference metric, which measures how far a predicted transparency matte departs from its reference. The comparison puts a 4B-parameter model’s fine-edge cutouts ahead of that larger, multi-model pipeline on this test—not on every image-matting task.
Tencent has released WeVisOne-VL in 2B and 4B sizes, alongside model weights, an inference package, evaluation code and a hosted demo. The project combines object detection, referring localization, counting, pointing and OCR with segmentation and alpha matting, allowing users to ask for either a box around an object or a pixel-level cutout.
A box can lead to a hair-level matte
The technical report, submitted to arXiv on October 22, describes WeVisOne-VL as a model built on Qwen3-VL. Its authors attach a 9.6-million-parameter pixel decoder to the vision-language model; that decoder uses the model’s image features and a special <OBJ> token to generate masks and continuous alpha mattes. The decoder is a component of the system, not the size of the full model.
For matting, users can identify a target with a box, a category name or a natural-language description. A binary mask marks pixels as either inside or outside an object. An alpha matte can represent partially transparent pixels, helping retain fine structures such as hair or fur and translucent boundaries. The model also uses dedicated coordinate tokens to produce points and boxes, and can return “None” when a requested object is absent.
The referring-matting result stands out
On Tencent’s WeRefMat test for targets specified by language, WeVisOne-VL-4B records a mean absolute difference of 9.8, against 22.4 for Qwen3.5-9B followed by SAM2-Matting. The 2B version scores 15.8 on that same comparison. The report also tests box-prompted matting on public datasets: on P3M-500-NP, WeVisOne-VL-2B records 6.6 and the 4B version 7.3, compared with 15.4 for the tested SAM2-Matting SAM3 variant.
Those numbers do not establish a universal ranking. The report says all methods on the three public box-prompted datasets receive the same ground-truth box, while the language-prompt comparison uses Tencent’s own WeRefMat benchmark. The researchers report mean absolute difference multiplied by 1,000, with lower values better; comparisons across different prompts or datasets would not be like-for-like.
One interface, several perception tasks
The release extends beyond cutouts. On LVIS, a long-tail object-detection benchmark, the team reports mean F1 scores of 53.3 for WeVisOne-VL-2B and 55.3 for the 4B model, compared with 50.7 for LocateAnything-3B. On Dense200, which tests detection in dense scenes, the corresponding scores are 61.4 and 71.3, versus 58.7 for that baseline. The report defines mean F1 as an average across intersection-over-union thresholds from 0.5 to 0.95.
WeVisOne-VL’s shared format treats tasks as structured text output. It can return object labels and coordinates, then attach a pixel-level result to a localized instance through the <OBJ> token. Tencent says the model was trained through three stages of supervised fine-tuning followed by mixed-task reinforcement learning, using a data engine built around roughly 100 million images.
The paper names limits on general use
“First, the limited amount of general-purpose training data may compromise the general capabilities relative to the base model,” the authors write in their limitations section. They also describe the current system as single-turn and say that omitting chain-of-thought reasoning for inference efficiency may limit complex reasoning. Clear benchmark prompts may not reflect vague or hesitant instructions from real users.
Tencent says its authored code is released under Apache 2.0, while third-party components retain their own licenses. The release is intended for research in visual grounding, OCR, segmentation and matting; the model card says it is not intended to identify individuals or make autonomous high-impact decisions.