AI Briefing
KOSign in

Tencent WeChat Vision WeVisOne-VL unifies object and pixel-level perception

·2026.10.23 17:14

Key point

Built on Qwen3-VL, WeVisOne-VL uses a 9.6M-parameter pixel decoder to unify object-level and pixel-level perception tasks.

Details

  • Tencent WeChat Vision developed WeVisOne-VL, a unified vision-language model built upon Qwen3-VL.
  • The model integrates object-level perception (open-vocabulary object detection, referring localization, object counting, object pointing, OCR) and pixel-level perception (referring segmentation and promptable alpha matting).
  • The architecture relies on structured next-token prediction and a lightweight 9.6M-parameter pixel decoder.
  • The model is available on Hugging Face, with the provided evidence linking to the tencent/WeVisOne-VL-2B repository.
vision-language-modeltencentobject-detectionsegmentationqwen3-vlhugging-face
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditNew LoRA for Qwen3-VL-8B Adds Bounding Box Coordinates to Document ExtractionRedditInterfaze 1 Lite Released: Mixture-of-Architectures Model for Developer WorkloadsHuggingFace BlogIBM releases Granite Speech 5.0 with improved ASR accuracy and 1.6x faster Turbo CTCHuggingFace BlogNVIDIA Releases SimReady Central: Open Library of 2,000+ Simulation-Ready 3D Assets on Hugging FacePyTorchKRHugging Face Provides Unified API for Local and Cloud LLMs with AnyLanguageModel 1.0PyTorchKRAWS TorchNeuron Runs PyTorch Code by Specifying Neuron Instead of CUDA