Tencent Releases WeVisOne-VL in 2B, 4B

WeVisOne-VL’s strongest idea is consolidation: one compact model handles tasks developers typically split across several specialized vision systems.

• –A lightweight 9.6M-parameter pixel decoder produces masks and alpha mattes without requiring SAM or an external segmentation backbone. [Official method overview](https://tencent.github.io/WeVisOne-VL/)
• –Structured coordinate tokens unify boxes, points, masks, OCR grounding, and absent-target rejection in one output format.
• –Tencent reports strong 4B results, including 55.3 LVIS F1, 96.4% rejection on WeRef-Bench, and 71.2 ReasonSeg gIoU. [Benchmark results](https://github.com/Tencent/WeVisOne-VL)
• –Apache 2.0 licensing, downloadable weights, evaluation code, and local inference make it practical for privacy-sensitive visual search, document processing, and image-editing pipelines.
• –The main caveat is deployment cost: “2B” and “4B” still require serious GPU resources for high-resolution pixel tasks, so specialists may remain cheaper for narrow workloads.