Alibaba Open Sources First Language Model Qwen-AgentWorld, Building Virtual Flight Simulator for AI Agents

According to Dynamic Insight Beating monitoring, TONGYI Lab has open-sourced the native language world model Qwen-AgentWorld. For the first time, the model takes environment modeling as a training target, training AI to predict the next environmental response, creating a virtual space for AI agents similar to a flight simulator. The virtual simulation avoids the high cost and security risks of agents trial-and-error in a real environment or network sandbox.

Qwen-AgentWorld covers a total of seven domains in a unified manner, including text and graphical interfaces. For graphical environments such as Web, OS, and Android, the model does not generate video frames but converts observations into code texts such as HTML and Accessible Tree XML, achieving ultra-fast and accurate logic simulation. In the comprehensive benchmark AgentWorldBench, Qwen-AgentWorld-397B-A17B achieves the highest overall score (58.71), surpassing GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro.

The model demonstrates two application values in agent training. On the one hand, as a decoupled environment simulator, the model can simulate thousands of unencountered virtual environments at zero cost, achieving training effects on par with or even surpassing real search engines in WideSearch tasks. On the other hand, the predictive capability can be internalized into the agent's meta-reasoning mode, enabling the same model to simulate environmental responses before action, resulting in significant gains in completely unencountered domains (Claw-Eval improvement +11.3, function call BFCL v4 improvement +9.0). The relevant models, benchmarks, and code have all been open-sourced.