Tencent releases WeVisOne-VL, a unified vision-language model for multiple AI tasks

Tencent's WeChat Vision team has introduced WeVisOne-VL, an open vision language model available in 2B and 4B parameter sizes. This perception-focused model, built upon Qwen3-VL, integrates several complex capabilities into a single architecture, including object detection, referring, counting, pointing, Optical Character Recognition (OCR), segmentation, and alpha matting. The model achieves this unification by casting every task as structured next-token prediction, utilizing coordinate tokens for normalized outputs. The architecture features a 9.6M-parameter decoder, eliminating the need for a separate Segment Anything Model (SAM). The training regimen involved three Supervised Fine-Tuning (SFT) stages: aligning output formats on 90M examples, increasing input capacity to 1,960 visual tokens and up to 500 boxes per response, and a final stage specializing the pixel decoder for matting on 130M samples. The 4B model generally outperforms the 2B model across most benchmarks, though the smaller version still achieves a 6.6 MAD on P3M-500-NP for box-prompted matting. Weights for both models are accessible on Hugging Face via the wevisone_vl package, which supports tasks like point and count_box. Tencent has also made its Mask2Alpha annotator open-source, while hosted demos for both model sizes are currently live.