tencent-WeVisOne-VL-4B-GGUF

WeVisOne-VL-4B (developed by WeChat Vision, Tencent Inc.) is an open-source, unified vision-language model built upon Qwen3-VL that bridges broad visual understanding with precise, multi-granularity spatial perception by integrating both object-level tasks (such as open-vocabulary object detection, referring localization, counting, pointing, and OCR) and pixel-level tasks (including referring segmentation and promptable alpha matting) through structured next-token prediction. Utilizing dedicated vocabulary tokens (<0> through <1000>) to represent normalized spatial coordinates and emitting an instance token (<OBJ>) after predicted bounding boxes, the model relies on a lightweight 9.6M-parameter pixel decoder that reuses multi-level ViT features via a visual feature adapter, a box-guided query decoder with deformable cross-attention, and an instance-conditioned dense prediction cascade (FullCanvasAlphaCascadeHead) to generate high-resolution continuous alpha mattes. Trained via a rigorous three-stage curriculum supervised fine-tuning and mixed-task reinforcement learning pipeline on approximately 100 million images, WeVisOne-VL achieves robust performance across diverse vision benchmarks while reliably handling target absence through explicit None rejections.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp