Qwen releases AgentWorld language world model
Qwen has released Qwen-AgentWorld, a single language model trained to simulate seven agent environments, which the team says can stand in for real environments during agent reinforcement learning.
Alibaba's Qwen team announced Qwen-AgentWorld, a "native language world model" trained to predict what an agent environment returns after the agent takes an action. A single model covers four text domains (MCP, Search, Terminal and SWE) and three GUI domains (Web, OS and Android). For the GUI domains, the model predicts renderable code such as accessibility-tree XML, HTML and UI hierarchy markup rather than pixel frames.
Qwen says environment modeling is the training objective at every stage, not something added to a general-purpose LLM afterward. The model goes through continual pre-training, supervised fine-tuning and reinforcement learning on more than 10 million real environment interaction trajectories:
- Pre-training: uses turn-level, information-theoretic loss masking to skip turns that carry little environment information.
- Fine-tuning: uses 7,094 rejection-sampled samples with explicit thinking traces.
- Reinforcement learning: uses GSPO, scored by a rubric-based LLM judge plus rule-based verifiers.
Qwen also released AgentWorldBench, a seven-domain benchmark that pairs predictions with ground-truth observations from real environments. Qwen says Qwen-AgentWorld-397B-A17B had the highest overall simulation quality on it, ahead of GPT-5.4, Claude Opus 4.8 and Gemini 3.1 Pro. The material provided here does not include the scores.
The team tested two uses for the model:
- As a simulator for agent RL: Qwen says training in the simulator "significantly outperforms" RL in real environments alone.
- As a warm-up for agent models: without RL on agentic tasks, it transferred to seven multi-turn agent benchmarks, three of them out of domain.
The model and benchmark are on Hugging Face and ModelScope, alongside a paper and GitHub repository.