Zikai分享Qwen-AgentWorld技术细节

Found one of the author of qwen-AgentWorld. Zikai shared some interesting insights: 1/ AgentWorld is beyond a simulator, one can definitely use it as an agent. 2/ 8b dense is the smallest model that can effectively learn from the simulation environment, 4b dense performs very bad, learns nothing from it. 3/ in practice the best input format for models trained using AgentWorld or AW itself would be accessibility tree given by playwright 4/ AgentWorld was trained on qwen3.5 not 3.6 bc the project started from Aug. 2025 when 3.6 not exist yet  > **引用原帖 Qwen (@Alibaba_Qwen):** > We open-source Qwen-AgentWorld-35B-A3B (MoE, 35B/3B active, 256K context) and AgentWorldBench. > Two routes, one roadmap: > 🔬 Build the simulator — scalable, controllable, surpassing real environments > 🧠 Internalize world modeling — predict before you act > Qwen-AgentWorld is our attempt to investigate how language world modeling can further expand the boundaries of general agent capabilities. > Go build on it 🏃🏃‍♂️ > 📑 Paper: https://t.co/Jx2l5RKq71 > 📖 Blog: https://t.co/7tVcKyhsx2 > 💻 GitHub: https://t.co/B5Lvb1UZCn > 🤗 HuggingFace: https://t.co/Kw3QBL1TM5 > 🧩 ModelScope: https://t.co/YBnGYgMWWI > https://x.com/Alibaba_Qwen/status/2069720412481888400