WeVisOne-VL-2B : nouveautés, matériel requis, comment l'utiliser gratuitement

Tencent released WeVisOne-VL-2B, a 2.4 billion-parameter vision-language model that handles objects and pixels in one go. It runs on a modest RTX 4060 with just 4 GB memory.

WeVisOne-VL-2B is a vision-language model from Tencent's WeChat Vision team. According to the model card, it builds on Qwen3-VL and adds a lightweight 9.5 million-parameter pixel decoder. The model can predict bounding boxes, points, and full-resolution alpha mattes directly from images, using special coordinate and instance tokens.

The new part is a dedicated pixel decoder that combines a visual feature adapter, a box-guided query decoder, and an instance-conditioned dense prediction head. It produces dense outputs like segmentation and matting without a separate processing stage. The model also returns None when a queried target is absent, simplifying downstream handling.

According to the model card, the model excels at open-vocabulary object detection, referring localization, counting, OCR, segmentation, and promptable alpha matting. However, fine alpha boundaries can be affected by image resolution, severe occlusion, complex transparency, and how prompts are phrased. It is designed for single-turn image perception; video grids are rejected.

The release is aimed at research on visual grounding, referring localization, pointing, counting, OCR localization, segmentation, and image matting. Because it fits in the memory of an RTX 4060 at 4 GB, hobbyists with a modest GPU can also experiment. No free API is available yet, but the weights can be downloaded from Hugging Face.

Que faut-il pour le faire tourner chez vous ? En 4 bits, il prend environ 4 gigaoctets de mémoire. La plus petite machine personnelle qui le contient entièrement est RTX 4060. Sur RTX 3060 12 GB avec 32 gigaoctets de RAM, il écrit environ 105 tokens par seconde. Sur RTX 4090 avec 64 gigaoctets de RAM, il écrit environ 247 tokens par seconde.

Il n’existe pas encore d’API gratuite pour cela. Les poids sont ouverts : vous pouvez les télécharger depuis Hugging Face sous licence Apache 2.0. Arena ne l’a pas encore classé : nous ajouterons le score dès que les votes arriveront.

Compare WeVisOne VL 2B with every free model, and check what your own GPU can run, on FreeLLMArena.