WeChat open-sources a 2B/4B vision model, and the comments first focus on whether it dares to say “None”

The WeChat Vision team announced WeVisOne-VL on X, an open-source 2B/4B vision model. The official tweet gives its capabilities very directly: finding objects, counting, reading text, generating precise masks, and making cutouts, without separately hooking up SAM. The repository is here, and the original tweet is here.

SAM is the usual kind of segmentation model, and for fine-grained segmentation you often have to attach a separate one. WeVisOne-VL puts detection, counting, text reading, masks, and cutout into a single 2B/4B model. Its selling point is not parameter scale but “one fewer component to wire in.” The tweet also specifically says: if the object is not in the image, output “None.”

The discussion under the tweet says more about the problem than the parameter table. One user said the first thing they wanted to test was this “None.” Small vision models have an old bad habit: when there is no target in the image, they still tend to force a box and are unwilling to admit the image is empty. Another user put it more directly: if a 2B model can hold its own on counting overlapping objects, then removing SAM would be a real gain.

Someone also laid out the rejection-rate gap. A user noted that the rejection rate is 96.4% on WeRef-Bench and 67.4% on RefCOCOg. The first number looks good; the second is the in-the-wild scenario. The model is very good at saying no to the “none” cases taught by the data engine, but whether it still works outside that distribution needs to be run. The same user also asked the flip side: when the object is clearly there, how often does it say “None”? This question is closer to real use than looking at rejection rate alone.

Someone also gave concrete advice for matting post-processing. A user suggested using task='matting' for background removal, keeping the continuous alpha output, and not thresholding it into a binary mask, otherwise fine edges will be lost. During QA, you can do a simple check: composite the same cutout onto a bright background and a dark background separately, and see whether hair strands and semi-transparent edges have halos.

There were two other concerns in the comments: not stuffing it into local WeChat, and whether edge devices have enough compute. These are unrelated to model capability and related to how it is deployed. Ordinary users care about whether the cutout is clean; developers care about rejection rate, overlapping-object counting, and edge quality. The advantage of open source is that you can run all of this yourself instead of just listening to a launch event. First test whether it force-draws a box on an empty image, then test whether it says “None” when an object is present, and finally look at cutout edges. The parameter count is not large, but the questions are very concrete.