Qwen drops AgentWorld foundation models for simulating agent environments

Training agents directly in live environments is expensive and unscalable, making Qwen's pivot to world modeling a necessary step for the next generation of autonomous systems. The models simulate environments across 7 domains, eliminating the need for real-world execution during RL training. By answering "what happens next" instead of "what should the agent do", Qwen-AgentWorld enables safe, massive-scale exploration of edge cases. The open-weights release includes massive 35B and 397B parameter models, alongside AgentWorldBench for evaluating simulation accuracy.