Qwen Releases 3 WebWorld Models

WebWorld is a series of open-web world models for web agents, fine-tuned based on Qwen3.

• Trained on 1M+ real-world web interaction trajectories.
• 32B/14B/8B checkpoints have been released, supporting 30+ step long-horizon simulation.
• State representation covers a wide range including A11y Tree, HTML, XML, Markdown, and natural language.
• It is explained that CoT-activated transition prediction generalizes to code, GUI, and game environments.
• Agents trained on WebWorld synthetic trajectories achieved MiniWob++ +9.9% and WebArena +10.9%.
• In inference-time lookahead search, it is presented as showing better world model performance than GPT-5.