Tencent has released WeVisOne-VL, an open vision-language model in 2B and 4B sizes that handles in one architecture tasks usually split across separate tools: object detection, referring localization, counting, pointing, OCR, segmentation and alpha matting, the extraction of a subject with soft edges.

Psst!Add AlexTech.ai asPreferred Source on Google

For developers the most concrete point is replacement: one model can stand in for a detector, an OCR engine and a segmenter. Pixel masks and mattes come not from an external SAM but from a 9.6-million-parameter decoder attached to the model. When the queried object is absent, the model answers "None" instead of inventing a box.

How it works

Every task is cast as structured next-token prediction. Normalized box and point coordinates are encoded as dedicated tokens. After each predicted box, the model emits a special token whose hidden state feeds the pixel decoder. The decoder has three parts: an adapter that reuses multi-level visual features from the Qwen3-VL base model, six box-guided attention blocks, and a head that refines the mask to full resolution.

Training and data

A knowledge-graph-driven data engine annotated over 100 million images with about 3.6 billion boxes. Candidate regions were proposed by SAM3, Rex-Omni and an in-house detector, and verified by Qwen3-VL. Training runs three supervised stages, followed by a final reinforcement stage on 178,000 uncertainty-filtered examples.

The reported numbers

The results are self-reported, using greedy decoding and at most 1,960 visual tokens per image. On Dense200 detection the 4B model scores 71.3 F1 mean, against 58.7 for LocateAnything-3B, a 3-billion-parameter model. On RefCOCOg referring it scores 81.8 against 77.6. On PixMo-Count exact-match accuracy is 89.6%. On TotalText OCR it scores 52.0 against 25.7 for the specialist PP-OCRv5. On ReasonSeg segmentation the gIoU is 71.2, against 65.0 for LISA-13B. On P3M-500-NP matting the mean absolute error is 7.3, where lower is better, against 15.4 for SAM2-Matting.

official image of WeVisOne-VL

Image: Tencent — https://github.com/Tencent/WeVisOne-VL

The weak point is rejection of absent targets: on PR-Bench the model correctly refuses only 52.6% of the time, so many queries about nonexistent objects still get a box. The 2B model trails the 4B on most tasks.

Against rivals

In the comparison with Qwen3-VL, Rex-Omni and LocateAnything, WeVisOne-VL-4B is the only one in the table with native masks and mattes, and it ships under Apache 2.0, unlike the non-commercial license of LocateAnything-3B.

How to use it

Weights for both sizes are on Hugging Face, and the wevisone_vl package supports Transformers and vLLM backends. Tencent has also published hosted demos and the matting annotator from its data engine. Fine-tuning code is still listed as to-do. GGUF builds cannot produce masks, because the decoder needs native visual features.

For computer vision teams the signal is clear: a lightweight decoder makes a separate segmenter unnecessary, and an open license makes it usable in commercial projects.