Member-only story
For the entire reinforcement-learning era, we have trained agents to act. Press the right button: run this command, click that element, call this tool. Qwen’s new release does the opposite. It trains a model to be the thing the agent acts on. You hand Qwen-AgentWorld a state (a terminal, a web page, an Android screen) and an action, and it predicts what comes back: the exact stdout, the next page of HTML, the JSON an API would return. It is a world model, but one that speaks text instead of pixels. On Qwen’s own benchmark for this, the 397B version scores 58.71, ahead of GPT-5.4 (58.25), Claude Opus 4.8 (56.59), and Gemini 3.1 Pro (54.57) at predicting what an environment will do. That number is the least interesting thing in the paper. The interesting thing is what becomes possible once a model can convincingly pretend to be the world: you can train agents inside a simulation you control completely, and, stranger still, there is early evidence that the act of learning to predict the world makes the model a better agent in it.
Key takeaways #
Qwen-AgentWorld(Qwen Team, Alibaba,arXiv:2606.24597, 23 June 2026) is a family of “language world models”: models trained to simulate agentic environments by predicting the next state as text, across seven domains (MCP…