Tencent Opens 31.7B WeLM-Image With Dual-Path MoT for Unified Generation and Editing

Tencent's WeChat Vision released WeLM-Image, a 31.7B-parameter model that splits into a 2.2B-active understanding path and a 5.4B-active generation path via a dual-path Mixture-of-Transformers plus 64 routed experts per path. The paper reports leading open-source scores on PRISM-Bench and Qwen-Image-Bench, with 27 edit action types and multi-reference synthesis. A WeLM-Image-Turbo variant renders 1024x1024 images in 4.98 seconds on an H20 after 8-step distillation.