Qwen-AgentWorld

What is Qwen-AgentWorld?

Qwen-AgentWorld is the industry’s first “native language” system, officially launched by the Tongyi Qianwen team on June 24, 2026.world model”(Language World Model, LWM). It is not a traditional general-purpose large language model, but rather a foundational model specifically designed and trained for AI agents. Based on over 10 million real-world interaction trajectories, the model is built through a three-stage training process—CPT (Continuous Pre-training) → SFT (Supervised Fine-tuning) → RL (Reinforcement Learning). It is designed to enable agents to mentally “simulate” and predict environmental feedback before performing real-world actions, thereby making smarter and safer decisions. Currently, the model has been fully open-sourced and is available in two sizes: 35B-A3B (lightweight version) and 397B-A17B (flagship version).

Key Features of Qwen-AgentWorld

1. Simulation of a Unified Environment Across the Seven Realms: A single model can cover seven major agent interaction environments. These include text-based environments (MCP tool invocation, Search, Terminal, SWE software engineering) and GUI-based environments (Web browser, OS operating system, Android mobile devices). In the GUI domain, the model represents states using renderable code—such as accessibility trees and HTML—rather than raw pixels, thereby enabling the modeling of visual environments using plain text.
2. State Prediction Using Long Thought Chains: The model takes the current state and the agent's action as input and, through Chain-of-Thought (CoT) reasoning, accurately predicts environmental feedback such as the next screen of the UI, device output, or error messages.
3. Controlled Adversarial Simulation: Supports the injection of specific simulation commands via natural language instructions (such as “hide some search results” or “simulate a disk full error”), systematically generating edge cases that are rare in real-world environments for use in adversarial training of agents.
4. Cross-Domain Knowledge Transfer: Thanks to native world modeling, knowledge acquired by the model in one domain can be transferred to another; for example, the skills learned in “navigating web pages” can help it learn “using mobile apps” more quickly.

Use Cases for Qwen-AgentWorld

The core value of Qwen-AgentWorld lies in providing AI agents with a low-cost, highly controllable “virtual training ground” and decision-making support. It is primarily used in the following scenarios:

1. Reinforcement Learning Training for Agents (Decoupled Simulator): By replacing real-world environments or sandboxes, AI agents can undergo large-scale, turn-based reinforcement learning training in a virtual environment. This approach is not only cost-effective and fast, but it also prevents irreversible dangerous actions that might occur in real-world environments.
2. Decision Enhancement for Agents (Unified Basic Model): By serving directly as the agent’s backbone model, the ability to model the environment is internalized as a forward-thinking mode characterized by “reflection” or “think-then-act.” Before actually executing an action, the agent first predicts environmental changes internally, thereby improving performance on downstream tasks.
3. Enterprise-Level Automation and Digital Employees: Suitable for complex tasks that require seamless switching between the command line, web pages, desktop software, and mobile apps—such as intelligent customer service, office automation, and software testing—it can significantly reduce the difficulty and cost of multi-agent collaborative development.

Project Address

• GitHub Repository: https://github.com/QwenLM/Qwen-AgentWorld
• ModelScope open-source repository: https://modelscope.cn/collections/Qwen/qwen-agentworld
• Hugging Face: https://huggingface.co/collections/Qwen/qwen-agentworld
• Demo URL: https://qwen.ai/blog?id=qwen-agentworld#interactive-demo-interactive-demo

How do I use Qwen-AgentWorld?

Qwen-AgentWorld supports two complementary usage paradigms:

1. As a Decoupled Environment Simulator: Developers can deploy the model as a standalone environment simulator and receive the current (state, action), the model will return the predicted next_state. This approach enables large-scale reinforcement learning training for agents without the need to set up actual sandboxes or virtual machines.
2. As a Unified Foundation Model for Intelligent Agents (Unified): Qwen-AgentWorld serves as the core backbone of the agent. Its built-in world modeling capabilities assist in action selection, and it is ready to use out of the box on benchmarks such as Terminal-Bench, SWE-Bench, and Claw-Eval, demonstrating strong generalization capabilities without requiring additional RL fine-tuning for specific agent tasks.

Comparison of similar products