Alibaba's Tongyi Lab has open-sourced Qwen-AgentWorld, a language world model that treats environment modeling as a training objective for the first time. By training AI to predict how an environment will respond next, it creates a virtual space — essentially a flight simulator for AI agents. This avoids the high costs and security risks of trial-and-error in real environments or network sandboxes.
Qwen-AgentWorld covers seven domains spanning text and graphical interfaces. For graphical environments like the web, desktop OS, and Android, the model doesn't generate video frames. Instead, it converts observations into code text such as HTML and accessibility tree XML, enabling ultrafast and precise logical simulation. On the comprehensive AgentWorldBench benchmark, the Qwen-AgentWorld-397B-A17B model achieved the highest overall average score (58.71), surpassing GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro.
The model demonstrates two practical applications for agent training. First, as a decoupled environment simulator, it can simulate thousands of unseen virtual environments at zero cost — in WideSearch tasks, it matched or even outperformed training with real search engines. Second, its predictive ability can be internalized as a meta-reasoning mode, allowing the same model to simulate environment responses before acting. This led to significant gains in completely unseen domains: +11.3 on Claw-Eval and +9.0 on function-calling benchmark BFCL v4. The models, benchmarks, and code are all open-sourced.