WeVisOne-VL Benchmarks: Where Tencent's Biggest Wins Come From

Ask most vision models for "the person on the left" and they'll draw a box. Ask for a person who isn't in the photo and many will draw one anyway. WeVisOne-VL, an open model from WeChat's vision team at Tencent, is built around answering that second question with a plain None, and its WeVisOne-VL benchmarks are the story here.

The team's account on X announced the release on Nov. 3, 2026, twelve days after the paper and weights went public. The pitch is one model for jobs that usually take three: detection, segmentation and matting. The paper and repository show where that holds up and where the headline numbers lean on the team's own tests.

WeChat's announcement on X, Nov. 3, 2026 (@Weixin_WeChat)

View on X

The post reads: "we gave our AI an eye for detail. Meet WeVisOne-VL, an open-source 2B/4B model from our Weixin Vision team. It finds and counts objects, reads text, and makes precise masks and cutouts—no separate SAM needed. Not in the image? 'None.'"

What ships, and when

The arXiv paper was submitted Oct. 22, 2026 and revised Oct. 24. The Hugging Face model page shows the 4B weights created that same day. Two sizes are public, 2B and 4B, and the 4B model card says it is built on Qwen3-VL. The GitHub README lists the tasks as detection, referring, counting, pointing, OCR, segmentation and matting, all driven by one predict call:

``python result = model.predict(image, task="matting", query="person") ``

The masks come from a 9.6-million-parameter pixel decoder bolted onto the language model, so the "no SAM needed" line is true at inference. It's less true of training. The paper says the labels came from a data engine that gathers box proposals from SAM3 and an in-house detector, then has a vision-language model verify them. SAM isn't a dependency of the model. It's part of its ancestry.

The table the README leads with

Tencent's headline table compares the 4B model against "the strongest open baseline in each benchmark." That's a different rival on every row. The paper's own text is more specific about who built what, and that changes how to read the gaps:

The paper describes WeRef-Bench as "an in-house benchmark comprising 2,880 English and Chinese queries" and uses a second in-house set, WeRefMat, for matting from a text prompt. Those are the two biggest wins in the table: 23.4 points on rejecting absent targets, and a 56% lower matting error (our arithmetic: 22.4 to 9.8). Public tests still show real gains, including 53% lower error on P3M-500-NP (15.4 to 7.3, ours again). But the largest gaps sit on tests only the authors could design around their own model.

The WeRefMat rival deserves a note too. It isn't a released product. The paper says it's a pipeline the authors assembled: Qwen3.5-9B finds a box, then SAM2-Matting cuts it out.

Smaller print in the full tables

Open the README's expandable results and the picture gets less uniform:

• On COCO detection, the 4B model scores 57.1 against 56.6 for GroundingDINO-T, a specialist detector. A half-point lead.
• On box-prompted matting, the 2B model beats the 4B on P3M-500-NP (6.6 against 7.3) and on PPM-100 (5.3 against 6.1).
• On pointing, LocateAnything-3B leads the 4B model on Dense200 (87.6 against 87.0).
• On RefCOCOg rejection, rivals such as Rex-Omni and LocateAnything score 0.5% and 0.6%. Those models weren't built around refusing, so a model trained to say None winning there tells you less than the 96.4% headline suggests. Qwen3.5-9B already reaches 70.7% on WeRef-Bench.

Our reading: the model is clearly strong, and the matting and counting gains on public data look real. The "one model replaces three" claim is the part to test yourself, with your own images.

What the model can't do

The model card says it "currently targets single-turn image perception" and that video grids are rejected by the pixel decoder. It also warns that fine alpha edges stay "sensitive to image resolution, severe occlusion, complex transparency, and prompt phrasing." The paper evaluates with greedy decoding and at most 1,960 visual tokens per image, which may matter for dense documents or very large photos. Code for fine-tuning is still unchecked on the README's to-do list.

Four answers before you try it

Is it free to use? The code is Apache 2.0 and the model page lists Apache 2.0, but third-party components keep their own licenses, as the card says.

How much GPU does it need? Unconfirmed. Neither the README nor the card gives a memory or speed figure.

Does it work on video? Not as released: the card says video grids are rejected.

Is the None answer reliable? On the team's tests, yes. Unconfirmed on anyone else's images.

A twenty-image test for the None answer

If you run detection or cutouts in a product, the claim worth checking is the rejection. Collect twenty images from your own pipeline where the target object is definitely absent, ask for it by the exact phrase your users type, and count how many come back None. Do the same in the languages you serve, since the team's own test mixes only English and Chinese. That count, not the 96.4% from a benchmark you can't inspect, is the number you'd ship against.

Source: Tencent, WeVisOne-VL repository

wevisone vl tencent wechat vision language model image matting

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.