Qwen-AgentWorld - The native language world model launched by Tongyi Qianwen

What is Qwen-AgentWorld

Qwen-AgentWorld is the first language world model launched by the Tongyi Qianwen team, simulating 7 major agent environments—MCP, search, terminal, software engineering, Android, Web, and operating system—through long chain-of-thought reasoning. The model is built on over 10 million real interaction trajectories and trained through a three-stage pipeline (CPT→SFT→RL), and the AgentWorldBench evaluation benchmark has also been launched to verify its leading performance in environment simulation and agent training.

Main features of Qwen-AgentWorld

• Unified seven-domain environment simulation: Covers 7 major interaction environments—MCP tool calling, search engines, Linux terminals, software engineering (SWE), Android GUI, Web browsers, and operating systems—using a unified text format to represent state transitions.
• Long chain-of-thought state prediction: Given the current state and agent action, the model predicts precise environment feedback through long CoT reasoning, such as the next screen UI, terminal output, error messages, etc.
• Controllable adversarial simulation: Supports injecting specific simulation instructions, such as "hide some search results" and "simulate a disk full error", to systematically generate edge cases that are rare in real environments.
• 4. Agent reinforcement learning training: Can serve as a decoupled environment simulator to support Sim Agentic RL, enabling scalable episode-level training in 4000+ real OpenClaw environments.
• Unified agent foundation model: World model training can serve as a "warm-up" for agent training, internalizing a forward-thinking mode similar to "reflection" to improve downstream task performance.

Technical principles of Qwen-AgentWorld

• Unified trajectory schema: Unifies the state representations of 7 heterogeneous environments into (system_prompt, action, observation) sequences, where system_prompt contains five parts: task description, action space, initial state, demonstration examples, and simulation instructions.
• Three-stage training pipeline：
  • CPT (continual pre-training): Injects state transition dynamics and enhanced specialized corpora to establish general world modeling capabilities.
  • SFT (supervised fine-tuning): Activates the reasoning chain-of-thought mode for "next-state prediction".
  • RL (reinforcement learning): Uses hybrid rubric-based rewards and rule-based rewards to refine simulation fidelity.
• Hybrid reward framework: For verifiable deterministic capabilities, such as terminal command execution and file system changes, rule verifiers are designed; for open-domain simulation quality, a five-dimensional rubric is used for evaluation.
• Environment representation strategy: The text domain directly predicts text output; the GUI domain represents states using accessibility trees and UI view hierarchies rather than raw pixels.

Follow WeChat and reply "open source" to joinAI open-source project exchange group

How to use Qwen-AgentWorld

• As an environment simulator (Decoupled): Deploy the model to replace the real environment, receive (state, action) through the API and return next_state, for large-scale episode-level RL training without a sandbox or virtual machine.
• As an agent foundation model (Unified): Using Qwen-AgentWorld directly as the agent backbone, its built-in world modeling capability can assist action selection, and it works out of the box on benchmarks such as Terminal-Bench, SWE-Bench, and Claw-Eval.
• Model acquisition: Download the weights from HuggingFace / ModelScope (Qwen-AgentWorld-35B-A3B , etc.)