Skip to main content
Back

Tencent's WeVisOne-VL Handles Seven Vision Tasks With one Open Model

Tencent's WeChat Vision team released an open 2B/4B vision-language model that handles detection, OCR, segmentation and alpha matting without a separate SAM backbone.

  • Tencent's WeChat Vision team open-sourced WeVisOne-VL, a 2B/4B VLM unifying object and pixel perception.
  • A 9.6M-parameter pixel decoder produces masks and alpha mattes with no SAM dependency.
  • 4B model leads on LVIS (55.3 F1), ReasonSeg (71.2 gIoU), and matting benchmarks.
  • Returns None for absent targets, hitting 96.4% rejection on WeRef-Bench vs 73% baseline.
  • Handles detection, referring, counting, pointing, OCR, segmentation and matting in one model.
  • Apache 2.0 license, weights on Hugging Face, with vLLM backend and Gradio demos.

Tencent’s WeChat Vision team has released the WeVisOne-VL repository, an open vision-language model that handles detection, referring, counting, pointing, OCR, segmentation, and matting through one set of weights. The release includes 2B and 4B parameter variants on Hugging Face under the Apache 2.0 license.

A 9.6M-parameter pixel decoder converts the model’s outputs into binary masks and alpha mattes, which represent per-pixel opacity. The released architecture generates these outputs without a SAM dependency, reducing the number of models required for applications that combine visual grounding with segmentation or subject extraction.

Seven tasks share one grammar

WeVisOne-VL expresses each perception task as structured text generation. Tokens from <0> through <1000> encode quantized box and point coordinates, None represents an absent target, and an <OBJ> token follows each box that requires a pixel-level output.

For segmentation and matting, the decoder combines the hidden state associated with each <OBJ> token, its predicted box, and multi-level features from the Qwen3-VL vision transformer. It then produces a full-resolution binary mask or alpha matte.

Structured output formats used by WeVisOne-VL across perception tasks
Boxes, points, null responses, masks, and mattes use a shared structured-output format.

For referring prompts with no matching object, the model can generate None. Explicit null output gives downstream systems a defined failure state instead of forcing every prompt to produce coordinates.

Reported gains concentrate in rejection and matting

Tencent compares the 4B model with Qwen3-VL models up to 9B parameters, Rex-Omni, LocateAnything, and specialist segmentation and matting systems. The project reports the following results: