Qwen-AgentWorld: Language World Models for General Agents

Researchers have developed Qwen-AgentWorld, the first large language-based world models capable of simulating agentic environments across 7 domains by predicting environment dynamics and next states. Trained on over 10 million real-world interaction trajectories using a three-stage pipeline, these models outperform existing frontier models and can be deployed either as scalable environment simulators for reinforcement learning or as foundation models that improve performance across multiple agent benchmarks.