Qwen-AgentWorld (arXiv paper) is a language world model created to predict what happens next in computer environments when actions are performed. A language world model is an artificial intelligence system trained to forecast future states or responses in interactive settings instead of deciding on actions. It handles seven domains in one system. These include text based areas but also visual interface areas such as web browsers, mobile operating systems, and desktop environments. For the visual areas the model works with text descriptions of screen elements rather than images. The model comes in two sizes. One is a mixture of experts design with thirty five billion total parameters but only about three billion active at once to reduce computing needs. The larger version has three hundred ninety seven billion total parameters.
Training approach and new evaluation method
The developers built the model through three connected training stages. First comes continual pre training to add broad knowledge about environments from large collections of interaction records. Next is supervised fine tuning where the model practices explicit step by step reasoning inside special thinking sections. Last is reinforcement learning where outputs receive scores from both automated rules and another model acting as judge to improve accuracy and realism. Along with the model the developers released AgentWorldBench. This is a test set built from real executions of other advanced models across nine earlier benchmarks. It measures prediction quality on five aspects including correctness of facts, logical consistency, and how realistic the outputs appear. On this test the larger model reached the highest overall score among compared systems and performed especially well on terminal and software editing tasks. The smaller model also showed strong results after the full training process.
The model supports two main uses for creating better agents. It can act as a separate practice environment that allows safe and controllable training at large scale because simulations can be adjusted in specific ways not possible in real systems. It can also serve as a base component inside agent systems so that agents learn to anticipate results before acting. This approach showed transfer benefits to new tasks without extra training on those specific tasks. The model and the test data are available for download from public repositories along with code for running evaluations and examples of use.