Alibaba's Qwen team has officially launched Qwen-AgentWorld, described as the first native language world model (LWM). The announcement was made on June 24, 2026. The model is designed to handle both text-based environments—such as MCP, Search, Terminal, and SWE—and GUI-based environments including Web, OS, and Android, all within a single framework.
Cross-domain knowledge transfer
Qwen-AgentWorld achieves cross-domain knowledge transfer by unifying diverse interaction paradigms. Unlike previous models that specialized in either text or graphical interfaces, this model can seamlessly switch between command-line tools and graphical user interfaces. This allows it to apply knowledge learned in one domain to tasks in another, potentially reducing the need for separate models for different types of software environments.
New benchmark for evaluation
Alongside the model, Qwen released AgentWorldBench, a benchmark for evaluating language world models across seven major domains. Each test sample in the benchmark includes real environment observations obtained by executing actions in actual environments, rather than relying on simulated data. This provides a more realistic assessment of model performance in practical scenarios.
The release positions Qwen-AgentWorld as a step toward more generalist AI agents that can operate across the full spectrum of digital interfaces, from code terminals to mobile apps.