Qwen-AgentWorld: A Language World Model for General-Purpose Agents

The Qwen team has unveiled Qwen-AgentWorld, a native language world model that simulates agent environments. Rather than simply post-adapting an LLM, this model makes environment modeling a direct training objective through a CPT (Continual Pre-training) → SFT → RL training pipeline.

Qwen-AgentWorld simulates the following 7 domains with a single model:

• Text-based: Terminal, Search, MCP (API), SWE (IDE)
• GUI-based: Web, Android, OS

In particular, for GUI environments, instead of pixel frames, the model implements text-based world modeling using renderable code such as Accessibility Tree XML, HTML. On the AgentWorldBench benchmark, the Qwen-AgentWorld-397B-A17B model demonstrated simulation quality surpassing GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro.

This model contributes to improving agent performance in two ways:

1. Decoupled environment simulator: Provides high scalability and controllability for agent reinforcement learning (RL).
2. Unified agent foundation model: Enables effective transfer learning to diverse multi-turn agent tasks even without separate RL fine-tuning.