Vision-language models have gotten fluent at describing images, but fine-grained perception still trips them up: counting every bird in a frame, tracing an exact contour, or cutting a foreground subject into a transparent alpha matte. Most models either hedge or confidently invent results for targets that are not even there. On October 22, the WeChat Vision Team at Tencent published a technical report on arXiv and open-sourced WeVisOne-VL, an attempt to pack object-level and pixel-level perception into one model behind a single interface.

One model, seven jobs

WeVisOne-VL stands for WeChat Vision All-in-One. Per the official README, it covers object detection, referring localization, counting, pointing, OCR, referring segmentation, and alpha matting, with output granularity spanning points, boxes, binary masks, and continuous alpha channels. The release ships 2B and 4B weight checkpoints, an inference package, cookbooks, and a full evaluation pipeline; the team also open-sourced Mask2Alpha, the annotator its data engine used to turn binary masks into high-quality alpha mattes. The GitHub repo was created on October 22 and sat at roughly 47 stars at writing time.

A 9.6M-parameter pixel decoder

The architectural story is lightness rather than scale: the model unifies tasks through structured next-token prediction, then a 9.6M-parameter pixel decoder plugs in via a special token to emit instance-level pixel outputs — no external SAM required. Training uses a three-stage curriculum SFT followed by mixed-task RL, and the team reports leading performance across more than 30 benchmarks at both object and pixel level.

The vendor-reported numbers

In the README's comparison table, WeVisOne-VL-4B faces the strongest open baseline per benchmark, including Qwen3-VL / Qwen3.5 models up to 9B, Rex-Omni, LocateAnything, and task-specific segmentation and matting models. The numbers: LVIS detection F1 55.3 vs 50.7; Dense200 dense detection 71.3 vs 58.7; RefCOCOg referring 81.8 vs 77.6; PixMo-Count accuracy 89.6 vs 82.8; TotalText OCR 52.0 vs 43.3; ReasonSeg referring segmentation 71.2 vs 65.0; P3M-500-NP box matting MAD 7.3 vs 15.4; WeRefMat text matting MAD 9.8 vs 22.4. The standout is rejection: on WeRef-Bench, the model refuses absent targets 96.4% of the time versus 73.0 for the baseline — the direct payoff of the "reliable rejection, no hallucination" design.

The cold water

Every number above is vendor-reported, and WeRef-Bench and WeRefMat carry the team's own "We" prefix — self-built benchmarks awaiting independent replication. The repo's TODO still lists the Gradio web demo and fine-tuning code as unreleased, and the license reads as custom ("Other"), so commercial use warrants a careful read. The traction is real, though: the paper topped the Hugging Face Daily Papers board the day after release with 220 upvotes.

So what

More than bigger models or longer contexts, industrial deployment needs perception that dares to say "nothing here" — a 96.4% rejection rate matters more than any single benchmark gain. The 2B/4B-backbone-plus-9.6M-decoder recipe is another data point for splitting perception into cheap, dedicated components. The next question: when perception becomes a cheap special-purpose part, do general VLMs still need to grow pixel-level hands of their own?

Paper: arxiv.org/abs/2610.31329 — Code: github.com/Tencent/WeVisOne-VL