Qwen-AgentWorld - A native language world model launched by Tongyi Qianwen - AiBoss

What is Qwen-AgentWorld?

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

Main functions of Qwen-AgentWorld

Seven-Domain Unified Environment SimulationIt covers seven major interactive environments: MCP tool calls, search engines, Linux terminals, software engineering (SWE), Android GUI, web browsers, and operating systems, and uses a unified text format to represent state transitions.
Long thought chain state predictionInput the current state and the agent's actions, and the model will predict accurate environmental feedback through long CoT inference, such as the next screen UI, terminal output, error messages, etc.
Controllable combat simulationIt supports injecting specific simulation commands, such as "hide some search results" and "simulate disk full error", systematically generating edge cases that are rare in real environments.
4. Agent reinforcement learning trainingIt can serve as a decoupled environment simulator, supporting Sim Agentic RL and enabling scalable round-based training in 4000+ real OpenClaw environments.
Unified Agent Basic ModelWorld model training can serve as a "warm-up" for agent training, internalizing into a forward thinking mode similar to "reflection" to improve the performance of downstream tasks.

Qwen-AgentWorld's Technical Principles

Unified trajectory modeUnify the state representation of the 7 heterogeneous environments into (system_prompt, action, observation) The sequence, system_prompt, consists of five parts: task description, action space, initial state, demonstration example, and simulation instructions.
Three-stage training pipeline:
- CPT (Continuous Pre-training)Injecting state transition dynamics and enhancing specialized corpora to establish general world modeling capabilities.
- SFT (Supervised Fine-tuning): Activate the reasoning chain mode of "predicting the next state".
- RL (Reinforcement Learning)The simulation fidelity is refined by using a hybrid rubric-based and rule-based scoring system.
Hybrid reward frameworkFor verifiable deterministic capabilities, such as terminal command execution and file system changes, a rule-based verifier is designed; a five-dimensional scoring standard is used to evaluate the quality of open domain simulation.
Environment Representation StrategyText fields directly predict text output; GUI fields use accessibility trees and UI view hierarchies to represent state, rather than raw pixels.

How to use Qwen-AgentWorld

As an environment simulator (Decoupled)Deploy a model to replace the real environment, receiving data via API. (state, action) return next_stateIt is used for large-scale round-based RL training without the need for sandboxes or virtual machines.
As a foundational model for intelligent agents (Unified)Qwen-AgentWorld can be used directly as the backbone of the intelligent agent. Its built-in world modeling capabilities can 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.
Model AcquisitionDownload weights for HuggingFace / ModelScope (Qwen-AgentWorld-35B-A3B wait)

Qwen-AgentWorld's core advantages

The first native multi-domain language world modelStarting from the CPT stage, environment modeling will be trained end-to-end as an explicit goal, rather than as a post-hoc adaptation of a general large language model.
Significantly outperforms the leading baselineOn AgentWorldBench, version 397B-A17B (58.8 points) outperforms Claude Opus 4.8 (56.6), GPT-5.4 (58.2), Gemini 3.1 Pro (54.6), DeepSeek-V4-Pro (53.0) and Qwen3.6-Plus (50.8).
Scalable and controllable It can scale environments at the round-level without requiring real infrastructure and supports precise perturbation to generate adversarial examples; its controllable simulated RL significantly outperforms RL trained only in real environments.
Cross-domain generalization and warm-up migrationLWM training serves as agent warm-up and can be transferred to 7 benchmarks (3 of which are not present in the training set at all), demonstrating strong generalization without agent task RL fine-tuning.
Plain text overlay visual environmentThe GUI domain represents the state using accessibility tree/HTML/UI hierarchical tags, and the visual interaction environment can be covered by plain text world modeling.

Qwen-AgentWorld project address

Project official websitehttps://qwen.ai/blog?id=qwen-agentworld
GitHub repository: https://github.com/QwenLM/Qwen-AgentWorld
HuggingFace model libraryhttps://huggingface.co/collections/Qwen/qwen-agentworld
arXiv technical paper: https://arxiv.org/pdf/2606.24597

Comparison of Qwen-AgentWorld with similar competing products

Dimension	Qwen-AgentWorld	WebWorld
position	Seven-Domain Unified Language World Model (Native LWM)	World Model for Large-Scale Web Agent Training
Environmental Coverage	7 major domains: MCP, Search, Terminal, SWE, Web, Android, OS	Single domain: Web browser environment
State representation	Unified Text Mode: Accessibility Tree XML, HTML, UI Hierarchical Markup, Shell Output, Code Execution Results	Multiple format states: A11y tree, HTML, XML, Markdown, natural language (5 formats are adaptively preserved)
Training methods	CPT → SFT → RL three-stage end-to-end native training, with 10 million+ real trajectories.	Two-stage course: Extensive pre-training (raw Web dynamics) + Causal reasoning activation (explicit state transition reasoning)
Model size	35B-A3B / 397B-A17B	14B / 32B
Context window	Supports long contexts (the exact length is not specified, but it covers multiple rounds of long trajectories).	256K tokens
Simulation depth	Long-chain thinking (CoT) inference predicts the next state, supporting controlled adversarial simulation.	Supports multi-round simulation with 30+ consecutive steps and consistent state tracking.
Action space	Native action formats for each domain (Shell commands, tool calls, UI operations, etc.)	Unified Python style function calls (`click(bid)`,`fill(bid, text)`,`goto(url)` wait)
Evaluation Criteria	Self-built AgentWorldBench (2170 samples / 7 domains / 5-dimensional scoring + rule validation)	Web agent benchmarking based on WebArena, Mind2Web, and other platforms.
Core advantages	Cross-domain unification, seven-domain native architecture, controllable adversarial capabilities, Sim Agentic RL + unified basic model dual paradigm	Web domain specialization, multi-format compatibility, action space standardization, 256K long context.

Application Scenarios of Qwen-AgentWorld

Intelligent agent training infrastructureIt provides a low-cost, scalable, and controllable virtual training ground for AI agents, replacing expensive sandboxes and real API calls.
Edge Cases and Adversarial TestingGenerate rare error states in real-world environments, such as network timeouts, permission denials, and insufficient resources, to test the robustness of the agent.
Software Engineering AidSimulates code execution, test feedback, and Git operation results to help developers anticipate the consequences of their actions.
Automated UI TestingSimulates Android/Web/desktop application interactions, predicts page state changes after clicks/inputs, and is used for automated test script generation.
Tool usage and the MCP ecosystemSimulates MCP server responses and multiple toolchain orchestration, helping developers debug complex call logic without deploying real services.